Performance contract library building method and system based on large language model

Abstract

The invention provides a performance contract library building method and system based on a large language model. The method comprises the following steps: obtaining a business text or contract data and processing the business text or contract data to obtain contract semantic features; generating a contract adaptation degree containing a primary adaptation degree and a secondary adaptation degree based on the features; the target performance contract constructs contract association information according to the primary adaptation degree and the core index association strength and sends the contract association information to the reference performance contract; determining a preferential fusion contract through feature integrating degree calculation and double threshold screening according to the performance contract; and the target performance contract fuses the contract semantic features and the adaptation degrees of the two parties to generate fusion features, and the fusion features are analyzed and then integrated into a contract library. According to the method, efficient construction of the contract library is realized through multi-level feature processing and dynamic priority judgment. According to the method, the construction efficiency and matching precision of the performance contract library can be improved, and the retrieval efficiency and precision of the key information of the contract performance library are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and natural language processing technology, and more specifically, to a method and system for building a performance contract library based on a large language model. Background Art

[0002] In today's business environment, enterprises face the need to manage massive amounts of contract data, which is not only massive but also complex and diverse. Traditional contract management methods rely primarily on manual review and classification, which is inefficient and prone to errors. With the advancement of information technology, some enterprises have begun to explore the use of electronic document management systems to store and retrieve contracts. However, most of these systems only provide basic document storage and keyword search functions, and are unable to provide in-depth semantic understanding and analysis of contract content.

[0003] The rapid development of artificial intelligence (AI) technology in recent years, particularly the emergence of large language models (LLMs), has brought new opportunities to contract management. LLMs can semantically understand text, extract key information, and generate natural language descriptions. However, current LLM-based contract management technologies primarily focus on the review and analysis of individual contracts, lacking comprehensive assessment of the relevance and compatibility between multiple contracts. Furthermore, existing technologies also have shortcomings in contract data processing and feature extraction, failing to effectively transform contract semantic features into structured data suitable for integration and matching.

[0004] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the existing technology: the existing technology cannot effectively process and fuse the semantic features between multiple contracts, making it difficult to achieve accurate matching and association between contracts; in the calculation of contract adaptability, there is a lack of comprehensive consideration of the correlation strength of key attributes and elements, resulting in inaccurate matching results; in addition, in the process of processing contract data, the existing technology cannot effectively extract and enhance the semantic features of the contract, which affects the efficiency and quality of subsequent fusion and matching. Summary of the Invention

[0005] The present invention provides a method and system for building a performance contract library based on a large language model.

[0006] In a first aspect of the present invention, a method for building a performance contract library based on a large language model is provided, comprising:

[0007] Obtain the corresponding business text or contract data;

[0008] Processing the business text or contract data to obtain contract semantic features;

[0009] Processing the semantic features of the contract to obtain contract suitability, wherein the contract suitability includes a primary suitability corresponding to the target performance contract and a secondary suitability corresponding to the reference performance contract;

[0010] The target performance contract generates contract association information based on the primary fit and the strength of association between the core indicators of the contract semantic features and the elements in the primary fit, and sends the performance contract association information to the reference performance contract;

[0011] Determining a contract for priority fusion information based on the contract-related information and the secondary fitness with reference to the performance contract, including: calculating a feature fit between the contract-related information and the corresponding secondary fitness with reference to the performance contract to obtain first data; performing a priority determination on each element in the first data based on a first threshold to obtain second data, wherein the priority determination includes: if an element exceeds the first threshold, setting the corresponding data point to a high priority, and setting the remaining elements to a low priority; performing contract screening on the second data according to a second threshold to determine a contract for priority fusion information; the contract screening includes: determining whether the second data exceeds the second threshold, and if so, selecting the corresponding contract as the contract for priority fusion information;

[0012] The target performance contract receives the contract semantic features and the corresponding secondary fitness of the contract from the priority fusion information, and fuses the contract semantic features of the priority fusion information and the target performance contract according to the primary fitness and the secondary fitness to obtain a fusion feature;

[0013] The target performance contract analyzes the fusion features to obtain the matched contract information, and integrates the matched contract information into the performance contract library to complete the construction of the performance contract library.

[0014] Furthermore, the business text is processed to obtain contract semantic features, including: segmenting the business text to obtain multiple text units; encoding the business text in the text unit to obtain a semantic vector; performing feature extraction and integration on the semantic vector to obtain a structured feature matrix; using a feature enhancement operator to process the structured feature matrix to obtain preliminary feature information; and performing feature extraction on the preliminary feature information to obtain contract semantic features.

[0015] Furthermore, the contract data is processed to obtain contract semantic features, including: using a semantic understanding module to predict the semantic category corresponding to each clause in the contract data, obtaining a semantic label, and mapping the contract data to a semantic space based on the semantic label; transforming the semantic label based on a contract element template to obtain element features; and combining the element features to obtain contract semantic features.

[0016] Furthermore, the contract semantic features are processed to obtain the contract adaptability, including: performing feature mapping on the contract semantic features to obtain multiple secondary feature information of different dimensions; performing reverse mapping on the multiple secondary feature information to obtain tertiary feature information of unified dimensions; splicing the multiple tertiary feature information in sequence, and performing irrelevant feature elimination processing on the spliced ​​feature information to obtain fourth-level feature information; matching contract elements based on the fourth-level feature information to obtain the contract adaptability.

[0017] Furthermore, based on the primary fitness, the strength of the association between the key attributes of the contract semantic features and each element in the primary fitness, the contract association information is constructed, including: determining the strength of the association between the key attributes of the contract semantic features and each element in the primary fitness; determining the degree of association corresponding to the primary fitness, the calculation formula of which is:

[0018]

[0019] Among them, Rel i Indicates the degree of correlation between the target performance contract numbered i, KeyF i (k) represents the kth key attribute of the target performance contract numbered i, PrimA i (k) represents the association strength of the kth element in the primary suitability of the target performance contract numbered i, and N represents the total number of key attributes; the contract association information is obtained by performing a weighted sum calculation on the association strength and the association closeness.

[0020] Furthermore, the contract semantic features of the contract with priority fusion information and the target performance contract are fused based on the primary fitness and the secondary fitness to obtain fusion features, including: interactively calculating the primary fitness and the secondary fitness through an associated weight mechanism to obtain auxiliary data; and multiplying the secondary fitness with the corresponding auxiliary data to obtain a fusion coefficient; utilizing a feature combination function to perform a dot product on the contract semantic features of the contract with priority fusion information and the corresponding fusion coefficient, and superimposing the dot product result with the contract semantic features of the target performance contract to obtain a fusion feature.

[0021] Furthermore, the target performance contract parses the fusion features to obtain matched contract information, including: the target performance contract parses the fusion features in sequence at the basic feature layer, enhanced feature extraction layer, prediction layer, and key element screening layer to obtain matched contract information, wherein the contract information includes clause categories and amount ranges.

[0022] Furthermore, referring to the performance contract, based on the contract association information and the secondary adaptability, a contract with priority fusion information is determined, and its expression is:

[0023]

[0024] Among them, D i,j A represents the degree of fit between the reference performance contract information numbered j and the target performance contract information numbered i. i (x,y) represents the contract association information of the target performance contract at position (x,y), B j (x, y) represents the secondary fitness of the reference performance contract at position (x, y), H represents the dimension of the contract semantic feature, and W represents the length of the contract semantic feature.

[0025] Furthermore, the expression for obtaining the fusion feature is:

[0026]

[0027] Among them, F i represents the fusion feature vector, FC represents the feature combination function, β i,j represents the fusion coefficient, C j Indicates the semantic features of the contract with reference to the performance contract, C i Represents the contract semantic features of the target performance contract. This expression is used to fuse the contract semantic features of the target performance contract and the contract with priority fusion information to generate a fused feature vector.

[0028] In a second aspect of the present invention, a system for building a performance contract library based on a large language model is provided, comprising:

[0029] Acquisition module, obtains the corresponding business text or contract data;

[0030] A first processing module processes the business text or contract data to obtain contract semantic features; the contract semantic features are processed to obtain contract suitability, wherein the contract suitability includes a primary suitability corresponding to the target performance contract and a secondary suitability corresponding to the reference performance contract;

[0031] The second processing module, the target performance contract receives the primary fitness from the first processing module, and constructs contract association information based on the primary fitness and the core indicators of the contract semantic features and the strength of the association between the elements in the primary fitness, and sends the performance contract association information to the reference performance contract;

[0032] The third processing module determines the contract for priority fusion information based on the secondary fitness from the first processing module and the contract association information from the second processing module with reference to the performance contract, including: calculating the feature fit of the contract association information and the corresponding secondary fitness with reference to the performance contract to obtain first data; performing priority determination on each element in the first data based on a first threshold to obtain second data, wherein the priority determination includes: if an element exceeds the first threshold, setting the corresponding data point to a high priority and the rest to a low priority; performing contract screening on the second data according to the second threshold to determine the contract for priority fusion information; the contract screening includes: determining whether the second data exceeds the second threshold, and if so, selecting the corresponding contract as the contract for priority fusion information;

[0033] In the detection module, the target performance contract receives the contract semantic features and the corresponding secondary fitness of the contract from the priority fusion information, and fuses the contract semantic features of the priority fusion information and the target performance contract based on the primary fitness and the secondary fitness to obtain the fusion features; the fusion features are parsed to obtain the matched contract information, and the matched contract information is integrated into the performance contract library to complete the construction of the performance contract library.

[0034] According to the above-mentioned embodiments of the present invention, there are at least the following beneficial effects: the performance contract library construction method and system based on the large language model of the present invention can achieve deep semantic understanding and efficient processing of contract data. By performing a series of operations such as segmentation, encoding, feature extraction and enhancement on business text or contract data, complex contract content can be converted into structured semantic features, thereby providing a solid foundation for subsequent contract fitness calculation and fusion. The method can also construct contract-related information based on the strength of the correlation between the key attributes of the contract semantic features and the elements in the fitness, and accurately determine the contracts that prioritize the fusion information through steps such as feature fit calculation, priority determination and contract screening, thereby effectively improving the accuracy and efficiency of contract fusion. In addition, through the analysis and matching of fusion features, the matched contract information can be integrated into the performance contract library, realizing the efficient construction of the performance contract library, and providing enterprises with a comprehensive, accurate and easy-to-manage contract data resource.

[0035] The present invention can also interactively calculate the primary fitness and the secondary fitness through an associated weight mechanism to obtain a fusion coefficient, and use a feature combination function to perform a dot product of the contract semantic features of the contract with priority fusion information and the fusion coefficient, and then superimpose them with the contract semantic features of the target performance contract to obtain a fusion feature. This process can not only make full use of the semantic information of the target performance contract and the reference performance contract, but also improve the representativeness and reliability of the fusion feature through weighted fusion. Ultimately, the target performance contract can perform multi-layer parsing processing on the fusion feature to obtain matched contract information including clause categories and amount ranges, which can further improve the accuracy and availability of the contract information and enhance the retrieval efficiency and accuracy of key information in the contract performance database. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which:

[0037] Figure 1 A flowchart of a method for building a performance contract database based on a large language model provided by one embodiment of the present invention;

[0038] Figure 2 A schematic diagram of the structure of a performance contract library building system based on a large language model provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0039] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. Rather, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0040] Those skilled in the art will appreciate that the embodiments of the present invention may be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present invention may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software.

[0041] It should be noted that any number of elements in the drawings is for illustration only and not for limitation, and any naming is only for distinction and does not have any limiting meaning.

[0042] Reference below Figure 1 , Figure 1This is a flow chart of a method for building a performance contract library based on a large language model according to an embodiment of the present invention. Figure 1 As shown, a method for building a performance contract library based on a large language model includes:

[0043] S1. Obtain the corresponding business text or contract data;

[0044] S2. Processing the business text or contract data to obtain contract semantic features;

[0045] S3. Processing the semantic features of the contract to obtain contract suitability, wherein the contract suitability includes a primary suitability corresponding to the target performance contract and a secondary suitability corresponding to the reference performance contract;

[0046] S4. The target performance contract generates contract association information based on the primary fit and the strength of association between the core indicators of the contract semantic features and the elements in the primary fit, and sends the performance contract association information to the reference performance contract;

[0047] S5. Determining, with reference to the performance contract, based on the contract-related information and the secondary fitness, contracts for priority fusion, including: calculating feature fit between the contract-related information and the corresponding secondary fitness with reference to the performance contract to obtain first data; determining the priority of each element in the first data based on a first threshold to obtain second data, wherein the priority determination includes: if an element exceeds the first threshold, setting the corresponding data point to a high priority and the remaining elements to a low priority; screening the second data for contracts according to a second threshold to determine contracts for priority fusion; the contract screening includes: determining whether the second data exceeds the second threshold, and if so, selecting the corresponding contract as the contract for priority fusion;

[0048] S6. The target performance contract receives the contract semantic features and the corresponding secondary fitness of the contract from the priority fusion information, and fuses the contract semantic features of the priority fusion information and the target performance contract according to the primary fitness and the secondary fitness to obtain a fusion feature.

[0049] S7. The target performance contract analyzes the fusion features to obtain the matched contract information, and integrates the matched contract information into the performance contract library to complete the construction of the performance contract library.

[0050] It should be noted that business text or contract data refers to various contract documents generated in the daily operations of an enterprise, including but not limited to procurement contracts, sales contracts, service contracts, etc. These documents usually exist in text form and contain rich business terms and key information. By processing these texts, contract semantic features that can reflect the essence of the contract content can be extracted, that is, feature vectors that can express the meaning of the contract terms in a digital way, thereby providing a basis for the contract's fitness evaluation and integration. In addition, the present invention also involves further processing of contract semantic features to obtain contract fitness. Fitness is a quantitative indicator that measures the similarity or degree of matching between contracts, including the primary fitness corresponding to the target performance contract and the secondary fitness corresponding to the reference performance contract. This helps to accurately select and integrate the reference contract that best matches the target contract when constructing a performance contract library.

[0051] Specifically, primary fit refers to the degree of match between the target performance contract and the reference performance contract in key business terms and core indicators. It reflects the similarity between the target and reference contracts in their primary business logic and key elements. Secondary fit, on the other hand, refers to the degree of match in non-core terms or secondary elements, and is used to assist in assessing the overall fit between the contracts. For example, primary fit can involve matching core terms such as contract amount, delivery deadline, and quality standards, while secondary fit can involve matching non-core terms such as payment method and liability for breach of contract. When processing business text, the text must first be segmented—that is, long text is broken down into multiple units. Each unit can be a sentence, a paragraph, or a clause—to facilitate subsequent encoding and feature extraction. The encoding process converts text units into a digital form that can be processed by computers. This is typically achieved through the embedding layer of a large language model, mapping the text into a high-dimensional semantic space to obtain a semantic vector for each unit. These semantic vectors are then subjected to feature extraction and integration to produce a structured feature matrix. This process may involve operations such as dimensionality reduction and normalization to ensure that the feature matrix accurately reflects the semantic information of the text. Finally, the structured feature matrix is ​​processed using the feature enhancement operator to further improve the distinguishability and expressiveness of the features, obtain preliminary feature information, and then obtain the final contract semantic features through feature extraction.

[0052] Preferably, when extracting contract semantic features, a pre-trained large language model, such as a Transformer-based model, can be used and fine-tuned to adapt it to the specific business text domain. Specifically, input parameters include the business text dataset, pre-trained model weights, learning rate, and number of training rounds. During the training process, the model learns how to map business text into a semantic space and extract effective semantic features. Contract fit can be calculated by constructing a feature mapping network to map contract semantic features to secondary feature information of different dimensions. Then, through inverse mapping, tertiary feature information with unified dimensions can be obtained. After concatenating these features and eliminating irrelevant features, contract elements can be matched based on the strength of the correlation between key attributes and elements to determine contract fit. For example, the strength of the correlation can be determined by calculating the correlation coefficient between the key attributes and the fit factors, while the closeness of the correlation can be calculated using a weighted summation approach. The weights can be set based on business needs and experience to ensure that the calculated results accurately reflect the actual degree of fit between contracts.

[0053] In some embodiments, business text is processed to obtain contract semantic features, including: segmenting the business text to obtain multiple text units; encoding the business text in the text unit to obtain a semantic vector; performing feature extraction and integration on the semantic vector to obtain a structured feature matrix; processing the structured feature matrix using a feature enhancement operator to obtain preliminary feature information; and performing feature extraction on the preliminary feature information to obtain contract semantic features.

[0054] It should be noted that in the present invention, when processing business text to obtain contract semantic features, the business text is first segmented, that is, the long text is broken down into multiple smaller text units to facilitate subsequent processing and analysis. These text units can be sentences, paragraphs, or clauses, depending on the text structure and business requirements. Next, the business text in each text unit is encoded and converted into a semantic vector. A semantic vector is a high-dimensional digital vector that can represent the semantic meaning of the text. The mapping from text to semantic vectors can be achieved through the encoder of a large language model. Then, these semantic vectors are subjected to feature extraction and integration to obtain a structured feature matrix. The feature matrix is ​​a two-dimensional array that can organize and arrange the key information in the semantic vector to facilitate subsequent processing. Finally, the structured feature matrix is ​​processed using a feature enhancement operator to obtain preliminary feature information, and further contract semantic features are extracted. The contract semantic features are processed feature vectors that can more accurately reflect the core semantic information of the contract text and provide a basis for subsequent contract fit calculation and fusion.

[0055] Specifically, business text segmentation involves dividing the original contract text into multiple small text units according to specific rules. For example, text can be segmented into sentence units based on sentence ending symbols (such as periods and semicolons), or into paragraph units based on paragraph structure. For text unit encoding, the embedding layer in a large language model can be used to map each text unit into a high-dimensional semantic space, resulting in a semantic vector. The dimensionality of the semantic vector is typically determined by the model architecture and can range from several hundred to several thousand dimensions. During the feature extraction and integration phase, models such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs) can be used to process the semantic vector, extract key feature information, and integrate it into a structured feature matrix. The rows of the feature matrix can represent different text units, and the columns can represent different feature dimensions. Finally, a feature enhancement operator can be a mathematical operation or a neural network module used to enhance key information in the feature matrix, for example, through normalization, standardization, or the addition of noise to improve feature robustness and discriminability.

[0056] Preferably, when encoding business text, a pre-trained large language model, such as the BERT or GPT series models, can be used. These models have been pre-trained on large amounts of text data and can effectively capture the semantic information of the text. In specific implementations, the business text can be input into the model's embedding layer to obtain a semantic vector for each text unit. For example, for a contract containing multiple clauses, each clause can be input into the model as a text unit to obtain the corresponding semantic vector for each clause. During the feature extraction and integration phase, a convolutional neural network can be used with multiple convolution kernels to extract feature information at different scales. For example, the convolution kernel sizes can be set to 3, 5, and 7 to extract local, medium-scale, and large-scale features, respectively. These features are then concatenated and integrated to form a structured feature matrix. During the feature enhancement phase, batch normalization techniques can be used to process the feature matrix to reduce internal covariate shift and improve model training efficiency and stability. Through these refined steps, the semantic features of the contract text can be more effectively extracted, providing high-quality input for subsequent contract fit calculation and fusion.

[0057] In some embodiments, contract data is processed to obtain contract semantic features, including: using a semantic understanding module to predict the semantic category corresponding to each clause in the contract data, obtaining a semantic label, and mapping the contract data to a semantic space based on the semantic label; transforming the semantic label based on a contract element template to obtain element features; and combining the element features to obtain contract semantic features.

[0058] It should be noted that in the process of processing contract data to obtain contract semantic features, the present invention first uses a semantic understanding module to predict the semantic category corresponding to each clause in the contract data. The semantic category here refers to the core business intention expressed by the clause, such as payment terms, delivery terms, breach of contract liability, etc. Through the semantic understanding module, the contract terms can be mapped to predefined semantic categories to obtain semantic labels. Then, based on the semantic labels, the contract data is mapped to the semantic space. The semantic space is a high-dimensional vector space that can represent the semantic information of the text in the form of vectors. Next, the semantic labels are transformed based on the contract element template to obtain element features. The contract element template is a predefined template used to extract key elements in the contract, such as amount, date, subject, etc. Finally, these element features are combined to obtain the contract semantic feature. The contract semantic feature is a feature vector that can comprehensively reflect the semantic information of the contract terms, providing a basis for subsequent contract adaptability calculation and fusion.

[0059] Specifically, the semantic understanding module is typically a trained deep learning model, such as one based on the Transformer architecture. It performs clause-by-clause semantic analysis on the contract text and predicts the semantic category of each clause. For example, for a clause requiring Party A to pay Party B 50% of the contract amount within 30 days of signing, the semantic understanding module can predict it as belonging to the semantic category of payment terms and assign it a corresponding semantic label. When mapping the contract data into a semantic space, word embedding technology can be used to convert each semantic label into a high-dimensional vector that maintains semantic similarity within the semantic space. The contract element template is a structured template that defines common elements in a contract and their extraction rules. For example, for the amount element, the template can define the extraction rule as the numeric portion of the contract amount, and for the date element, the template can define the extraction rule as the specific date format mentioned in the contract. Using these templates, key information from the semantic labels can be extracted to form element features. Finally, these element features are combined, for example, through vector concatenation or weighted summation, to obtain the final contract semantic features.

[0060] When building the semantic understanding module, a pre-trained language model such as BERT can be used and fine-tuned to adapt to the specific characteristics of contract text. During fine-tuning, the model can be trained using supervised learning, using contract terms as input and pre-defined semantic categories as labels. Input parameters include the contract text dataset, pre-trained model weights, learning rate, and number of training rounds. When mapping to the semantic space, word embedding techniques such as Word2Vec or GloVe can be used to convert each semantic label into a fixed-dimensional vector. Contract element templates can be designed based on actual business needs. For example, for a procurement contract, element templates such as purchase quantity and unit price can be defined. When extracting element features, regular expressions or named entity recognition techniques can be used to extract specific element values ​​from the text based on template rules. Finally, when combining element features, weights can be assigned based on their importance. For example, the amount element can be given a higher weight to reflect its importance in the contract. These refined steps allow for more accurate extraction of contract semantic features, providing high-quality input for subsequent contract fitness calculation and fusion.

[0061] In some embodiments, contract semantic features are processed to obtain contract adaptability, including: performing feature mapping on contract semantic features to obtain multiple secondary feature information of different dimensions; performing reverse mapping on multiple secondary feature information to obtain tertiary feature information of unified dimensions; sequentially splicing multiple tertiary feature information and performing irrelevant feature elimination processing on the spliced ​​feature information to obtain fourth-level feature information; and matching contract elements based on the fourth-level feature information to obtain contract adaptability.

[0062] It should be noted that in the process of processing the semantic features of the contract to obtain the contract adaptability, the present invention first performs feature mapping on the semantic features of the contract, that is, converting the original semantic features into secondary feature information of different dimensions. This process can be achieved through a neural network or other machine learning model. Then, the multiple secondary feature information are reverse mapped respectively, with the purpose of converting the feature information of different dimensions into tertiary feature information of unified dimensions for subsequent processing. Then, the multiple tertiary feature information are spliced ​​in sequence, and the spliced ​​feature information is subjected to irrelevant feature elimination processing to obtain fourth-level feature information. Finally, the contract elements are matched based on the fourth-level feature information to obtain the contract adaptability. The contract adaptability here refers to a quantitative indicator for measuring the similarity or matching degree between contracts, including the primary adaptability corresponding to the target performance contract and the secondary adaptability corresponding to the reference performance contract, which is used to evaluate the matching degree between contracts.

[0063] Specifically, feature mapping involves converting the original semantic features of a contract into new feature representations using a function or model. For example, a multi-layer perceptron (MLP) network can be used to map the input semantic feature vector into multiple dimensions, with each dimension corresponding to a second-level feature. These second-level features can reflect different aspects of the contract, such as clause type, amount range, and contract term. Inverse mapping involves converting these second-level features of different dimensions into third-level features of a unified dimension for subsequent concatenation. This process can be implemented using an inverse mapping function, such as a linear or nonlinear transformation to align features of different dimensions to a common dimension. Concatenation involves sequentially concatenating multiple third-level features to form a long feature vector. This concatenation may contain irrelevant features, necessitating the removal of irrelevant features. This can be achieved using a feature selection algorithm, such as one based on information gain or L1 regularization. Contract element matching involves calculating the similarity between contracts based on the fourth-level feature information to determine contract compatibility. This can be achieved by calculating the distance or similarity between feature vectors, such as using cosine similarity or Euclidean distance.

[0064] Preferably, when constructing the feature mapping model, a deep neural network, such as an MLP network containing multiple hidden layers, can be used. Input parameters include the original contract semantic feature vector and the network's structural parameters (such as the number of neurons in the hidden layer, activation function, etc.). During the inverse mapping process, a linear transformation matrix can be used to map secondary feature information of different dimensions to a unified dimension. For example, if the dimensions of the secondary feature information are 100, 150, and 200, respectively, they can be mapped to a unified dimension, such as 256, through a linear transformation. During the splicing and irrelevant feature elimination processing stage, a feature selection algorithm can be used, such as a feature selection method based on L1 regularization, which automatically eliminates unimportant features by training a linear model with L1 regularization. During the contract element matching stage, cosine similarity can be used to calculate the similarity between two contract feature vectors. Specifically, the dot product of the two vectors is divided by the product of their moduli to obtain a similarity value between 0 and 1. The closer the value is to 1, the higher the compatibility between the contracts. Through these detailed steps, the compatibility between contracts can be calculated more accurately, providing a reliable basis for subsequent contract integration and the construction of a performance contract library.

[0065] In some embodiments, based on the primary fitness, the strength of association between the key attributes of the contract semantic features and each element in the primary fitness, the contract association information is constructed, including: determining the strength of association between the key attributes of the contract semantic features and each element in the primary fitness; determining the degree of association corresponding to the primary fitness, which is calculated as follows:

[0066]

[0067] Among them, Rel i Indicates the degree of correlation between the target performance contract numbered i, KeyF i (k) represents the kth key attribute of the target performance contract numbered i, PrimA i (k) represents the association strength of the kth element in the primary suitability of the target performance contract numbered i, and N represents the total number of key attributes; the contract association information is obtained by performing a weighted sum calculation on the association strength and the association closeness.

[0068] It should be noted that key attributes refer to those semantic features of contracts that have a significant impact on contract matching and integration, such as contract value, delivery deadline, and quality standards. Association strength is a quantitative indicator that measures the correlation between these key attributes and the elements of primary fit. By combining association strength and association closeness through a weighted summation, contract association information is derived. This process effectively reflects the degree of semantic correlation between the target performance contract and the reference performance contract, providing an important reference for subsequent contract integration.

[0069] For example, in a procurement contract, key attributes may include purchase amount, delivery date, and supplier qualifications. These attributes are often closely related to the contract's core business logic. Primary fit refers to the degree of match between the core terms of the target performance contract and the reference performance contract, reflecting the primary similarity between the contracts. For example, if the purchase amount and delivery date of two contracts are very similar, their primary fit will be high. Association strength is calculated by calculating the correlation between key attributes and each element of primary fit. For example, the Pearson correlation coefficient can be used to measure the linear correlation between two variables. When calculating the closeness of association, the association strengths of all key attributes are weighted and summed. The weights can be set based on business needs and experience. For example, for procurement contracts, the amount can be given a higher weight, as it is often a key factor in contract matching. The resulting contract association information is a comprehensive indicator reflecting the degree of semantic association between the target contract and the reference contract, which can be used to guide contract integration and matching.

[0070] Furthermore, when calculating association strength, a pre-trained deep learning model can be used to assess the correlation between key attributes and the primary fitness factor. For example, a model based on the Transformer architecture can be used, with the input parameters including the feature vector of the key attribute and the feature vector of the primary fitness factor. The model calculates the correlation between the two feature vectors using a multi-head self-attention mechanism and outputs an association strength value. To determine the degree of association, the association strength values ​​of all key attributes can be weighted and summed. The weights can be adjusted based on business needs. For example, for a procurement contract focused on cost control, the weight of the amount attribute can be set to 0.5, the weight of the delivery date attribute to 0.3, and the weight of the supplier qualification attribute to 0.2. This ensures that the calculated association strength accurately reflects the actual semantic connection between the contracts. The resulting contract association information can be used to guide contract fusion and matching. For example, when the degree of association is high, the corresponding reference contract can be prioritized for fusion with the target contract.

[0071] In some embodiments, the contract semantic features of the contract with priority fusion information and the target performance contract are fused based on the primary fitness and the secondary fitness to obtain a fusion feature, including: interactively calculating the primary fitness and the secondary fitness through an associated weight mechanism to obtain auxiliary data; and multiplying the secondary fitness with the corresponding auxiliary data to obtain a fusion coefficient; using a feature combination function to perform a dot product on the contract semantic features of the priority fusion information and the corresponding fusion coefficient, and superimposing the dot product result with the contract semantic features of the target performance contract to obtain a fusion feature.

[0072] It should be noted that when the present invention fuses the contract semantic features of the priority fusion information and the target performance contract, first, the primary fitness and the secondary fitness are interactively calculated through the association weight mechanism to obtain auxiliary data. The association weight mechanism here refers to an algorithm that dynamically adjusts the weight, which is used to calculate the weight of each contract feature based on the relative importance of the primary fitness and the secondary fitness. Then, the secondary fitness is multiplied by the corresponding auxiliary data to obtain the fusion coefficient. The fusion coefficient is a quantitative indicator used to measure the importance of each reference contract feature in the fusion process. Finally, the feature combination function is used to perform a dot product on the contract semantic features of the priority fusion information contract and the corresponding fusion coefficient, and the dot product result is superimposed on the contract semantic features of the target performance contract to obtain the fusion feature. This process can effectively integrate the semantic information of the target contract and the reference contract, generate more representative fusion features, and provide support for subsequent contract information matching.

[0073] Specifically, the association weighting mechanism is a dynamically adjusted weighting algorithm designed to assign appropriate weights to each contract feature based on the relative importance of primary and secondary fitness factors. For example, if primary fitness factors (such as contract amount and delivery deadline) are more important than secondary fitness factors (such as payment method and default liability), the primary fitness factor will receive a higher weight. Auxiliary data, intermediate data calculated through the association weighting mechanism, is used to adjust the weights of secondary fitness factors. The fusion coefficient is calculated by multiplying the secondary fitness factor with the auxiliary data. It reflects the importance of each reference contract feature in the fusion process. For example, if a reference contract scores highly on secondary fitness factors and also has high auxiliary data, its fusion coefficient will be larger, indicating that this contract feature is more important in the fusion process. The feature combination function is a mathematical function used to combine multiple feature vectors. For example, feature vectors can be combined using weighted summation or dot product methods. In the dot product operation, the corresponding elements of each feature vector are multiplied and then summed to produce a scalar value that reflects the similarity between the feature vectors. Finally, this scalar value is superimposed with the contract semantic features of the target performance contract to obtain the fusion feature.

[0074] Preferably, a deep learning model, such as a neural network based on an attention mechanism, can be used when constructing the association weight mechanism. Input parameters include feature vectors for the primary and secondary fitness metrics. The model dynamically adjusts the weights using the attention mechanism and outputs auxiliary data. For example, a multi-head self-attention mechanism can be used. Taking two feature vectors as input, the model outputs a weight vector representing the importance of each feature in the fusion process. To calculate the fusion coefficient, each element of the secondary fitness metric can be multiplied by the corresponding element of the auxiliary data to obtain the fusion coefficient. For example, if the secondary fitness metric is a vector containing multiple features, each with a corresponding weight, a product operation can be used to obtain the fusion coefficient for each feature. In the feature combination function, a dot product operation can be used to perform a dot product between the contract semantic features of the priority fusion contract and the fusion coefficient to obtain a scalar value. This scalar value is then weighted and summed with the contract semantic features of the target performance contract to obtain the final fused feature. For example, the feature vector of the target performance contract can be added to the scalar value to obtain the fused feature vector. Through these refined steps, the semantic features of the target contract and the reference contract can be more effectively fused to generate more representative and accurate fusion features, providing high-quality input for subsequent contract information matching.

[0075] In some embodiments, the target performance contract parses the fusion features to obtain matched contract information, including: the target performance contract parses the fusion features in sequence at the basic feature layer, enhanced feature extraction layer, prediction layer, and key element screening layer to obtain matched contract information, wherein the contract information includes clause categories and amount ranges.

[0076] It should be noted that the present invention adopts a multi-layer parsing method in the process of parsing the fused features to obtain the matched contract information. This method can gradually extract the key information in the fused features by processing the fused features in the basic feature layer, enhanced feature extraction layer, prediction layer and key element screening layer in sequence, and finally obtain the matched contract information. The matched contract information here refers to the contract feature information after processing and fusion, which can accurately reflect the core terms and key elements of the contract, such as the term category and amount range. Through this multi-layer parsing process, the accuracy and availability of contract information can be effectively improved, providing high-quality data support for the construction of the performance contract library.

[0077] Specifically, the basic feature layer performs preliminary processing on the fused features, extracting essential feature information from them. This layer typically includes normalization and dimensionality reduction of the fused feature vector to facilitate subsequent processing. For example, PCA (Principal Component Analysis) can be used to reduce the dimensionality of the feature vector to reduce computational complexity. The enhanced feature extraction layer further extracts and enhances key feature information based on the basic feature layer. This layer can use models such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs) to perform convolution or recurrent operations on the basic features to extract deeper features. The prediction layer predicts key contract elements based on this enhanced feature information. For example, a classifier can be used to predict the contract clause category, or a regression model can be used to predict the amount range. The key element screening layer selects the most important elements from the prediction results to form the final matched contract information. This layer can select the most important elements by setting thresholds or using ranking algorithms. For example, key elements can be screened based on the prediction confidence or importance score.

[0078] Preferably, when constructing the basic feature layer, a simple linear transformation model can be used, with the fused feature vector as input and the normalized basic feature vector as output. In the enhanced feature extraction layer, a convolutional neural network can be used, with the basic feature vector as input. Enhanced features can be extracted through convolutional and pooling layers. For example, multiple convolution kernels can be configured, each extracting features at a different scale. In the prediction layer, a deep learning model, such as a multi-layer perceptron (MLP), can be used, with the enhanced feature vector as input and the output being predictions for the clause category and amount range. For example, the model can be divided into two branches: one for classification and one for regression. In the key element screening layer, a confidence-based screening algorithm can be used. For example, for clause category predictions, categories with a confidence level above a certain threshold (e.g., 0.8) are selected as key elements; for amount range predictions, predictions within a reasonable range are selected as key elements. Through these refined steps, the fused features can be more accurately parsed, resulting in high-quality matched contract information and providing reliable data support for the construction of a performance contract database.

[0079] In some embodiments, the contract with priority fusion information is determined based on the contract association information and the secondary fitness with reference to the performance contract, and the expression thereof is:

[0080]

[0081] Among them, D i,j A represents the degree of fit between the reference performance contract information numbered j and the target performance contract information numbered i. i (x,y) represents the contract association information of the target performance contract at position (x,y), B j (x, y) represents the secondary fitness of the reference performance contract at position (x, y), H represents the dimension of the contract semantic feature, and W represents the length of the contract semantic feature.

[0082] It should be noted that this expression is used to measure the degree of fit between the target performance contract and the reference performance contract to determine the contracts that are prioritized for fusion. In the process of determining the contracts for which information is prioritized for fusion, the present invention calculates the degree of fit between the target performance contract and the reference performance contract to determine which reference contracts best match the target contract in terms of semantics and business logic. The degree of fit here is a quantitative indicator used to measure the similarity and degree of match between two contracts. Through this evaluation method, contracts that are prioritized for fusion can be effectively screened out, thereby improving the accuracy and efficiency of contract fusion.

[0083] Specifically, the target performance contract refers to the core contract that needs to be integrated with other contracts and serves as the benchmark contract during the integration process. The reference performance contract refers to other contracts used for comparison and integration with the target performance contract. These contracts typically have similar business backgrounds or clause structures. Contractual relevance information refers to information obtained by analyzing the semantic features and key attributes of the target performance contract, and is used to describe the degree of semantic relevance between the target contract and other contracts. Secondary fit refers to the degree of match between the reference performance contract in non-core terms or minor elements, reflecting the similarity between the reference contract and the target contract in non-critical areas. When calculating fit, the contractual relevance information of the target performance contract and the secondary fit of the reference performance contract are comprehensively evaluated. A specific calculation formula is used to generate a quantitative value that measures the degree of fit between the two. For example, the corresponding elements of the contractual relevance information and the secondary fit can be multiplied and summed to obtain a comprehensive fit value.

[0084] Preferably, a deep learning model, such as one based on a Transformer architecture, can be used to calculate fit. Input parameters include the contract-related information of the target performance contract and the secondary fit of the reference performance contract. The model calculates the similarity between the two using a multi-head self-attention mechanism and outputs a fit value. For example, the feature vectors of the target and reference contracts can be input into the model, which will output a fit value between 0 and 1, with values ​​closer to 1 indicating a higher degree of fit. When determining the contracts to prioritize for fusion, they can be sorted by fit value, with the contracts with the highest fit selected as the priority contracts. For example, a threshold value, such as 0.7, can be set, and only contracts with a fit above this threshold are selected for priority fusion. This refined approach allows for a more accurate assessment of the degree of fit between contracts, thereby improving the accuracy and efficiency of contract fusion.

[0085] In some embodiments, the expression for obtaining the fusion feature is:

[0086]

[0087] Among them, F i represents the fusion feature vector, FC represents the feature combination function, β i,j represents the fusion coefficient, C j Indicates the semantic features of the contract with reference to the performance contract, C i Represents the contract semantic features of the target performance contract. This expression is used to fuse the contract semantic features of the target performance contract and the contract with priority fusion information to generate a fused feature vector.

[0088] It should be noted that the present invention adopts a method based on feature combination function in the process of obtaining fusion features. By fusing the contract semantic features of the contract with priority fusion information with the contract semantic features of the target performance contract, a fusion feature vector is generated. The core of this process is to perform weighted fusion of features through feature combination function to ensure that the fused feature vector can more accurately reflect the semantic association and matching degree between the target contract and the reference contract. The fused feature vector is an important foundation for subsequent contract information matching and performance contract library construction, and it can provide more accurate data support for contract management.

[0089] Specifically, the feature combination function is a mathematical function used to combine multiple feature vectors. It can be a simple weighted summation or a more complex nonlinear combination method, such as dot product, convolution, etc. In the present invention, the role of the feature combination function is to fuse the contract semantic features of the contract with priority fusion information with the contract semantic features of the target performance contract. The fusion coefficient here is a weight factor used to adjust the importance of the contract features of the priority fusion information in the fusion process. The calculation of the fusion coefficient is based on the secondary fitness and auxiliary data, reflecting the contribution of the reference contract in the fusion process. By performing a dot product operation on the contract semantic features of the contract with priority fusion information and the fusion coefficient, and superimposing the result with the contract semantic features of the target performance contract, a fusion feature vector is finally generated. This process can be regarded as a feature enhancement process, and a more representative feature vector is obtained by fusing the semantic features of multiple contracts.

[0090] Preferably, a deep learning model, such as a multi-layer perceptron (MLP) or a model based on the Transformer architecture, can be used when constructing the feature combination function. The input parameters include the contract semantic features of the contract with priority fusion information and the contract semantic features of the target performance contract. The model dynamically adjusts the fusion coefficient by learning the relationship between the features. For example, an MLP model with an attention mechanism can be used. Two feature vectors are input and the model outputs a fusion coefficient vector. In the dot product operation, the contract semantic features of the contract with priority fusion information are multiplied element-wise by the fusion coefficient vector to produce a weighted feature vector. This weighted feature vector is then element-wise added to the contract semantic features of the target performance contract to produce the final fused feature vector. When training the model, supervised learning methods can be used, using the known fused feature vector as a label and optimizing the model parameters through backpropagation. This approach can more effectively fuse the semantic features of the target and reference contracts, generating a more representative and accurate fused feature vector, providing high-quality input for subsequent contract information matching and performance contract library construction.

[0091] The aforementioned embodiments of the present invention have the following beneficial effects: This method can improve the efficiency of building a performance contract database. It uses a large language model to extract semantic features from business text and contract data, combined with multi-level fitness calculations, to achieve precise matching and integration of contract information. Key attribute correlation strength analysis and a dynamic priority screening mechanism can optimize the generation and transmission of contract-related information, ensuring efficient interaction between target performance contracts and reference contracts, thereby improving the data integration quality of the contract database.

[0092] Furthermore, this method enhances the ability to parse contract semantic features, utilizing techniques such as feature enhancement, inverse mapping, and irrelevant feature elimination to improve the accuracy of contract element matching. Through fusion coefficient calculation and feature combination functions, the weights of semantic features across different contracts can be dynamically adjusted to generate more reasonable fusion results. Ultimately, through multi-layer parsing and key element screening, structured contract information, including clause categories and amount ranges, can be output, enhancing the efficiency and accuracy of retrieval of key information from the contract performance database.

[0093] like Figure 2 As shown, some embodiments provide a system for building a performance contract library based on a large language model, the system comprising:

[0094] Acquisition module 201, acquiring corresponding business text or contract data;

[0095] The first processing module 202 processes the business text or contract data to obtain contract semantic features; the contract semantic features are processed to obtain contract suitability, wherein the contract suitability includes a primary suitability corresponding to the target performance contract and a secondary suitability corresponding to the reference performance contract;

[0096] In the second processing module 203, the target performance contract receives the primary fitness from the first processing module, constructs contract association information based on the primary fitness and the strength of association between the core indicators of the contract semantic features and the elements in the primary fitness, and sends the performance contract association information to the reference performance contract;

[0097] The third processing module 204 determines the contract for priority fusion information based on the secondary fitness from the first processing module and the contract association information from the second processing module with reference to the performance contract, including: calculating the feature fit of the contract association information and the corresponding secondary fitness with reference to the performance contract to obtain first data; performing priority determination on each element in the first data based on a first threshold to obtain second data, wherein the priority determination includes: if an element exceeds the first threshold, setting the corresponding data point to a high priority and the remaining elements to a low priority; performing contract screening on the second data according to the second threshold to determine the contract for priority fusion information; the contract screening includes: determining whether the second data exceeds the second threshold, and if so, selecting the corresponding contract as the contract for priority fusion information;

[0098] Detection module 205, the target performance contract receives the contract semantic features and the corresponding secondary fitness of the contract from the priority fusion information, and fuses the contract semantic features of the priority fusion information and the target performance contract based on the primary fitness and the secondary fitness to obtain the fusion features; the fusion features are parsed to obtain the matched contract information, and the matched contract information is integrated into the performance contract library to complete the construction of the performance contract library.

[0099] It is understandable that the modules and references recorded in the performance contract library construction system based on the large language model Figure 1 The steps in the method for building a performance contract library based on a large language model correspond to each other. Therefore, the operations, features and beneficial effects described above for the method for building a performance contract library based on a large language model are also applicable to the performance contract library building system based on a large language model and the modules contained therein, and will not be repeated here.

[0100] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or a terminal device such as a computer, a server, a mobile phone, or a tablet.

[0101] The above descriptions are merely some preferred embodiments of the present invention and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also encompass other technical solutions formed by any combination of the above-mentioned technical features or their equivalents without departing from the above-mentioned inventive concept. For example, a technical solution formed by mutually replacing the above-mentioned features with (but not limited to) technical features having similar functions disclosed in the embodiments of the present invention.

Claims

1. A method for building a performance contract database based on a large language model, characterized in that: include: Obtain the corresponding business text or contract data; Processing the business text or contract data to obtain contract semantic features; Processing the semantic features of the contract to obtain contract suitability, wherein the contract suitability includes a primary suitability corresponding to the target performance contract and a secondary suitability corresponding to the reference performance contract; The target performance contract generates contract association information based on the primary fit and the strength of association between the core indicators of the contract semantic features and the elements in the primary fit, and sends the performance contract association information to the reference performance contract; Determining a contract for priority fusion information based on the contract-related information and the secondary fitness with reference to the performance contract, including: calculating a feature fit between the contract-related information and the corresponding secondary fitness with reference to the performance contract to obtain first data; performing a priority determination on each element in the first data based on a first threshold to obtain second data, wherein the priority determination includes: if an element exceeds the first threshold, setting the corresponding data point to a high priority, and setting the remaining elements to a low priority; performing contract screening on the second data according to a second threshold to determine a contract for priority fusion information; the contract screening includes: determining whether the second data exceeds the second threshold, and if so, selecting the corresponding contract as the contract for priority fusion information; The target performance contract receives the contract semantic features and the corresponding secondary fitness of the contract from the priority fusion information, and fuses the contract semantic features of the priority fusion information and the target performance contract according to the primary fitness and the secondary fitness to obtain a fusion feature; The target performance contract analyzes the fusion features to obtain the matched contract information, and integrates the matched contract information into the performance contract library to complete the construction of the performance contract library.

2. A method for building a performance contract database based on a large language model according to claim 1, characterized in that: The business text is processed to obtain the contract semantic features, including: Segment the business text into multiple text units; Encode the business text in the text unit to obtain a semantic vector; Extracting and integrating features of the semantic vectors to obtain a structured feature matrix; Processing the structured feature matrix using a feature enhancement operator to obtain preliminary feature information; Feature extraction is performed on the preliminary feature information to obtain contract semantic features.

3. The method for building a performance contract database based on a large language model according to claim 1, characterized in that: Contract data is processed to obtain contract semantic features, including: The semantic understanding module is used to predict the semantic category corresponding to each clause in the contract data, obtain the semantic label, and map the contract data into the semantic space based on the semantic label; Transform semantic labels based on contract element templates to obtain element features; The element features are combined to obtain the contract semantic features.

4. The method for building a performance contract database based on a large language model according to claim 1, characterized in that: The contract semantic features are processed to obtain the contract adaptability, including: Perform feature mapping on the contract semantic features to obtain multiple secondary feature information of different dimensions; Perform reverse mapping on multiple secondary feature information to obtain tertiary feature information with unified dimensions; Multiple third-level feature information are sequentially spliced ​​together, and irrelevant features are eliminated from the spliced ​​feature information to obtain fourth-level feature information; Contract elements are matched based on the four-level feature information to obtain the contract suitability.

5. The method for building a performance contract database based on a large language model according to claim 1, characterized in that: Based on the primary fit, the key attributes of the contract semantic features and the strength of the association between the elements in the primary fit, the contract association information is constructed, including: Determine the strength of the association between the key attributes of the contract semantic features and the elements in the primary fitness; determine the closeness of the association corresponding to the primary fitness, and the calculation formula is: Among them, Rel i Indicates the degree of correlation between the target performance contract numbered i, KeyF i (k) represents the kth key attribute of the target performance contract numbered i, PrimA i (k) represents the association strength of the kth element in the primary fitness of the target performance contract numbered i, and N represents the total number of key attributes; The contract association information is obtained by performing weighted sum calculation on the association strength and association closeness.

6. The method for building a performance contract database based on a large language model according to claim 1, characterized in that: Based on the primary fitness and secondary fitness, the semantic features of the contract with priority fusion information and the target performance contract are fused to obtain fusion features, including: The primary fitness and secondary fitness are interactively calculated through the association weight mechanism to obtain auxiliary data; The secondary fitness is multiplied by the corresponding auxiliary data to obtain the fusion coefficient; The feature combination function is used to perform dot product between the contract semantic features of the contract with priority fusion information and the corresponding fusion coefficient, and the dot product result is superimposed with the contract semantic features of the target performance contract to obtain the fusion feature.

7. The method for building a performance contract database based on a large language model according to claim 1, characterized in that: The target performance contract analyzes the fusion features to obtain matching contract information, including: The target performance contract analyzes and processes the fusion features in the basic feature layer, enhanced feature extraction layer, prediction layer and key element screening layer in sequence to obtain the matched contract information, which includes the clause category and amount range.

8. The method for building a performance contract database based on a large language model according to claim 1, characterized in that: The performance contract is referred to based on the contract-related information and the secondary adaptability to determine the contract with priority fusion information, and its expression is: Among them, D i,j A represents the degree of fit between the reference performance contract information numbered j and the target performance contract information numbered i. i (x,y) represents the contract association information of the target performance contract at position (x,y), B j (x, y) represents the secondary fitness of the reference performance contract at position (x, y), H represents the dimension of the contract semantic feature, and W represents the length of the contract semantic feature.

9. A method for building a performance contract database based on a large language model according to claim 8, characterized in that: The expression for obtaining fusion features is: Among them, F i represents the fusion feature vector, FC represents the feature combination function, β i,j represents the fusion coefficient, C j Indicates the semantic features of the contract with reference to the performance contract, C i represents the contract semantic features of the target performance contract, and j represents the number of the reference performance contract.

10. A performance contract database construction system based on a large language model, characterized in that: include: Acquisition module, obtains the corresponding business text or contract data; A first processing module processes the business text or contract data to obtain contract semantic features; Processing the semantic features of the contract to obtain contract suitability, wherein the contract suitability includes a primary suitability corresponding to the target performance contract and a secondary suitability corresponding to the reference performance contract; The second processing module, the target performance contract receives the primary fitness from the first processing module, and constructs contract association information based on the primary fitness and the core indicators of the contract semantic features and the strength of the association between the elements in the primary fitness, and sends the performance contract association information to the reference performance contract; The third processing module determines the contract for priority fusion information based on the secondary fitness from the first processing module and the contract association information from the second processing module with reference to the performance contract, including: calculating the feature fit of the contract association information and the corresponding secondary fitness with reference to the performance contract to obtain first data; performing priority determination on each element in the first data based on a first threshold to obtain second data, wherein the priority determination includes: if an element exceeds the first threshold, setting the corresponding data point to a high priority and the rest to a low priority; performing contract screening on the second data according to the second threshold to determine the contract for priority fusion information; the contract screening includes: determining whether the second data exceeds the second threshold, and if so, selecting the corresponding contract as the contract for priority fusion information; In the detection module, the target performance contract receives the contract semantic features and the corresponding secondary fitness of the contract from the priority fusion information, and fuses the contract semantic features of the priority fusion information and the target performance contract based on the primary fitness and the secondary fitness to obtain the fusion features; the fusion features are parsed to obtain the matched contract information, and the matched contract information is integrated into the performance contract library to complete the construction of the performance contract library.

Images