A high school procurement budget and execution linkage management method

Abstract

The present application relates to the technical field of data processing management, in particular to a university procurement budget and execution linkage management method.The present application extracts the structured attributes of budget subjects to generate a budget feature matrix;collects unstructured text and structured form data of procurement execution flow, inputs a pre-trained language model to extract an execution feature vector;constructs a cross-modal attention matching network based on a sliding time window, takes the budget feature matrix as a key vector and a value vector, and takes the execution feature vector as a query vector;introduces a procurement stage mask mechanism in forward propagation, dynamically adjusts the weight coefficients of different feature dimensions according to the procurement life cycle stage, outputs a matching probability distribution and establishes a correlation mapping.The present application eliminates the mismatch phenomenon caused by the misalignment of unstructured text and structured code feature space, and reduces data processing delay.

Description

Technical Field

[0001] This invention relates to the field of management data processing technology, specifically a method for linking university procurement budget and execution management. Background Technology

[0002] In university procurement budget and execution linkage systems, the association between procurement business data and budget items is generally handled using manually pre-configured keyword dictionaries or fixed field mapping tables. Specifically, the budget management layer pre-maintains structured fields for budget items in the database. Upon receiving contract texts or expense reports during the procurement execution phase, the system calls the pre-established mapping rule table to perform precise string matching between specific fields in the procurement request and fixed terms in the dictionary, or directly maps the serial number to the corresponding budget item code according to fixed encoding truncation rules. Based on the matching results, a foreign key association is established between the procurement execution serial record and the budget item code in the database, completing subsequent budget deductions and status changes.

[0003] The implementation of string matching and fixed field mapping based on fixed vocabulary presents a single core technical problem when dealing with unstructured text generated during the procurement execution phase. The expression of unstructured text such as procurement contract terms and invoice summaries is highly variable, while budget item codes are highly abstract structured data. The string matching rules based on fixed vocabulary cannot identify the implicit semantic relationships between unstructured text expressions and budget item names, resulting in an insurmountable semantic gap between unstructured procurement text data and structured budget codes. This leads to association failures or mismatches due to the misalignment of their feature spaces. Summary of the Invention

[0004] The purpose of this invention is to provide a method for linking university procurement budget and execution management, which can effectively solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for linking university procurement budget and execution management includes: extracting structured attribute data of university budget items and generating a budget feature matrix; Collect unstructured text data and structured form data of the procurement execution flow, input the unstructured text data and the structured form data into a pre-trained language model, and extract execution feature vectors; A cross-modal attention matching network based on a sliding time window is constructed, with the budget feature matrix used as the key vector and value vector of the cross-modal attention matching network, and the execution feature vector used as the query vector of the cross-modal attention matching network. During the forward propagation of the cross-modal attention matching network, a procurement stage masking mechanism is introduced. Based on the procurement lifecycle stage of the procurement execution flow, the weight coefficients of different feature dimensions in the cross-modal attention matching network are dynamically adjusted, and the matching probability distribution of the execution feature vector and the budget feature matrix is ​​output. Based on the matching probability distribution and the preset probability threshold, an association mapping is established between the procurement execution flow and the corresponding budget item.

[0006] Preferably, the step of extracting structured attribute data of university budget items and generating a budget feature matrix includes: parsing the hierarchical relationship between university budget items and constructing a tree topology diagram of budget items; Traverse the tree topology of the budget items and obtain the item category code, the management department code, and the funding source code of each budget item node as the structured attribute data; The subject category code, the department code, and the funding source code are concatenated to obtain the initial feature vector of each budget subject node; The graph convolutional network is used to perform feature aggregation calculation on the tree topology graph of the budget items and the initial feature vector, and the updated hidden state of each budget item node is used as the row vector in the budget feature matrix.

[0007] Preferably, the step of inputting the unstructured text data and the structured form data into a pre-trained language model to extract execution feature vectors includes: performing key-value pair serialization concatenation on the field names and field values ​​in the structured form data to obtain a structured sequence; The unstructured text data is segmented into words to obtain a text segmentation sequence; The structured sequence is concatenated with the text segmentation sequence, and a first type of segment identifier is assigned to the structured sequence and a second type of segment identifier is assigned to the text segmentation sequence in the concatenated sequence. The concatenated sequence carrying the first segment identifier and the second segment identifier is input into the pre-trained language model, and the feature vector corresponding to the first segment identifier in the last hidden layer of the pre-trained language model is extracted as the execution feature vector.

[0008] Preferably, the construction of the cross-modal attention matching network based on a sliding time window includes: obtaining the current procurement execution flow set and sorting it in descending order according to the generation timestamp of each flow in the procurement execution flow set; Using the current moment as the end point of the time window and a preset time length as the span of the time window, a subset of the target flow within the time window is extracted from the sorted procurement execution flow set. The execution feature vectors corresponding to the target pipeline subset are used to form a query vector matrix; The budget feature matrix is ​​divided into a key vector matrix and a value vector matrix that match the dimension of the query vector matrix. The key vector matrix, the value vector matrix, and the query vector matrix are then input into the cross-modal attention matching network.

[0009] Preferably, the introduction of the procurement stage masking mechanism includes: obtaining the procurement lifecycle stage label of the current procurement execution flow, wherein the procurement lifecycle stage label includes the demand declaration stage, the contract signing stage, and the fund payment stage; Construct a stage embedding vector lookup table with the same number of procurement lifecycle stage tags, and retrieve the corresponding stage vector from the stage embedding vector lookup table based on the procurement lifecycle stage tags; A linear fully connected mapping is performed on the stage vector to generate a mask tensor with the same dimension as the attention weight matrix of the cross-modal attention matching network. The mask tensor is then superimposed on the attention score matrix of the cross-modal attention matching network.

[0010] Preferably, the step of dynamically adjusting the weight coefficients of different feature dimensions in the cross-modal attention matching network and outputting the matching probability distribution between the execution feature vector and the budget feature matrix includes: calculating the dot product of the query vector and the key vector to obtain the original attention score matrix; The original attention score matrix is ​​added element-wise to the mask tensor to obtain the mask attention score matrix; Perform a normalized exponential operation on each row of the mask attention score matrix to generate the weight coefficients; The value vector is weighted and summed using the weighting coefficients to obtain a fused feature vector. The fused feature vector is input into a fully connected layer and a normalized exponential function, and the matching probability distribution of the execution feature vector corresponding to each budget item is output.

[0011] Preferably, the step of using a graph convolutional network to perform feature aggregation calculation on the budget item tree topology graph and the initial feature vector includes: for any central node in the budget item tree topology graph, obtaining a set of neighboring nodes of the central node within a preset number of hops; Extract the initial feature vector corresponding to each neighbor node in the neighbor node set, and multiply each initial feature vector by the corresponding edge weight coefficient; The initial feature vectors after multiplication are summed to obtain the neighbor aggregation feature of the central node; The initial feature vector of the central node is concatenated with the aggregated features of its neighbors, and the concatenation result is input into an activation function to obtain the updated hidden state of the central node. The edge weight coefficient is calculated by decaying based on the shortest path length between two connected nodes in the budget subject tree topology graph.

[0012] Preferably, before extracting the feature vector corresponding to the first class segment identifier in the last hidden layer of the pre-trained language model, the method further includes: obtaining a historical procurement execution flow sample set and the corresponding historical budget item sample labels, inputting the historical procurement execution flow sample set into the pre-trained language model, and extracting sample feature vectors; Using the sample feature vector as the anchor point, positive sample feature vectors belonging to the same historical budget item label are taken as positive examples, and negative sample feature vectors belonging to different historical budget item labels are taken as negative examples, and the contrastive learning loss function is calculated. The model parameters of the pre-trained language model are updated based on the gradient backpropagation of the contrastive learning loss function, wherein the contrastive learning loss function adopts the InfoNCE loss function scaled by the temperature coefficient.

[0013] Preferably, the time window span is set to a preset time length, and the target flow subset within the time window is extracted from the sorted procurement execution flow set, including: counting the number of historical flows in adjacent historical time windows before the current time. Calculate the difference between the historical transaction volume and the preset baseline transaction volume; If the difference is greater than zero, the preset time length is shortened by the proportion of the difference; if the difference is less than zero, the preset time length is extended by the proportion of the difference, thus obtaining the adjusted actual time window span. Based on the actual time window span, the target flow subset is extracted from the sorted procurement execution flow set to maintain the number of flow items in the target flow subset within a fixed range.

[0014] Preferably, performing a linear fully connected mapping on the stage vector to generate a mask tensor with the same dimension as the attention weight matrix of the cross-modal attention matching network includes: dividing the execution feature vector into a name dimension feature subspace, an amount dimension feature subspace, and a time dimension feature subspace; When the procurement lifecycle stage label is the demand declaration stage, the mapping value in the stage vector corresponding to the name dimension feature subspace is set to a positive number and the mapping value corresponding to the amount dimension feature subspace is set to a negative number through the linear fully connected mapping. When the procurement lifecycle stage label is the payment stage, the mapping value corresponding to the amount dimension feature subspace in the stage vector is set to a positive number through the linear fully connected mapping, and the mapping value corresponding to the name dimension feature subspace is set to a negative number. The mask tensor is generated based on the set mapping value.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. To address the semantic gap between unstructured procurement text and structured budget code, which leads to association failure, this paper constructs a tree-like topology by parsing the hierarchical relationships of budget items. A graph convolutional network is then used to aggregate the initial feature vectors of nodes in the tree-like topology to generate a budget feature matrix. Simultaneously, the unstructured text of the procurement execution flow is concatenated with the structured form and assigned corresponding segment identifiers. This concatenation is then input into a pre-trained language model to extract execution feature vectors, ensuring that both unstructured text and structured budget code are transformed into the same feature space. The budget feature matrix is ​​used as the key and value vectors, and the execution feature vector is used as the query vector. These are then input into a cross-modal attention matching network to calculate the matching probability distribution. An association mapping is established based on the probability distribution threshold, eliminating mismatches caused by feature space misalignment.

[0016] 2. In the forward propagation of the cross-modal attention matching network, the labels of the procurement lifecycle stages of the procurement execution flow are obtained. Stage vectors are retrieved through a stage embedding vector lookup table. A linear fully connected mapping is performed on the stage vectors to generate a mask tensor with the same dimensions as the attention weight matrix. This mask tensor is then superimposed on the attention score matrix. Combining the name, amount, and time dimension feature subspaces of the execution feature vectors, the mapping values ​​of the corresponding feature subspaces in the mask tensor are set to positive or negative numbers at different procurement lifecycle stages, dynamically adjusting the weight coefficients of different feature dimensions in the attention calculation. Simultaneously, the procurement execution flow set is sorted in descending order of generation timestamp. A subset of target flows within a sliding time window is extracted, with the current time as the endpoint. The window span is shortened or extended proportionally based on the difference between the number of historical flows in adjacent historical time windows and the preset baseline number of flows, maintaining the number of target flow subsets in the input calculation within a fixed range, thus reducing data processing latency. Attached Figure Description

[0017] Figure 1 This is the overall flowchart of the university procurement budget and execution linkage management method of the present invention; Figure 2 This is a flowchart of the process for generating the budget feature matrix according to the present invention; Figure 3 This is a flowchart of the feature vector extraction process of the present invention; Figure 4 This is a flowchart illustrating the construction of the cross-modal attention matching network of the present invention; Figure 5 The flowchart for introducing a procurement stage masking mechanism and weight adjustment in this invention is shown below; Figure 6 This is a flowchart illustrating the specific dimensional control process for generating the mask tensor according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] Please refer to Figure 1 This embodiment provides a method for linking university procurement budget and execution management. This method runs on the server side of the university's integrated financial management platform. The server side establishes a data communication connection with the university's budget management system, procurement management system, contract management system, and fund payment system through the university's local area network. The data communication process adopts an encrypted transmission protocol that complies with the information security standards of the education industry. All data collection and processing comply with the university's financial internal control management specifications.

[0020] Specifically, the university budget management system stores full data of annual budget items that have passed the final review and approval process within the university. This full data includes basic information such as budget item code, item name, item level, responsible department, funding source, budget amount, available balance, and item status. First, the full budget item data undergoes data cleaning to remove duplicate entries, invalid budget item data (those that have been cancelled or frozen). The remaining valid budget item data is then structured and parsed to extract structured attribute data for each valid budget item. This structured attribute data includes item category code, responsible department code, and funding source code. The item category code conforms to the expenditure economic classification code rules, the responsible department code is a unique code for secondary management departments uniformly compiled within the university, and the funding source code includes predefined unique codes for different funding types such as financial revenue, operating revenue, and other revenue. The extracted structured attribute data is numerically encoded. The subject category code, the responsible management department code, and the funding source code are converted into fixed-dimensional numerical vectors using one-hot encoding. The one-hot vector corresponding to the subject category code has a dimension of 64, the one-hot vector corresponding to the responsible management department code has a dimension of 32, and the one-hot vector corresponding to the funding source code has a dimension of 16. The three one-hot vectors are concatenated in a preset order to obtain the initial feature vector corresponding to each budget subject. The dimension of the initial feature vector is 112. The initial feature vectors corresponding to all valid budget subjects are arranged in the order of budget subject codes to generate a budget feature matrix with a dimension of N×112, where N is the total number of valid budget subjects.

[0021] Furthermore, unstructured text data and structured form data of the procurement execution flow are collected. The procurement execution flow data comes from all procurement business data generated in the procurement management system, contract management system, and fund payment system. The structured form data includes fixed field data from purchase requisition forms, contract filing forms, expense reports, and payment application forms. Fixed fields include procurement project name, procurement amount, applying department, applicant, procurement method, contract number, invoice number, payment time, etc. Unstructured text data includes text content without a fixed format, such as procurement requirement descriptions, contract clauses, names of goods or taxable labor services on invoices, expense reimbursement remarks, and technical parameters in contract attachments. The collected structured form data is validated for integrity, and invalid flow data with empty field values ​​or field formats that do not conform to preset rules is removed. The collected unstructured text data is preprocessed to remove garbled characters, special symbols, and meaningless whitespace characters. Traditional Chinese characters are converted to simplified Chinese, and English words are converted to lowercase.

[0022] The preprocessed unstructured text data and structured form data are input into the pre-trained language model to extract execution feature vectors. For each field name and its corresponding value in the structured form data, key-value pairs are concatenated. All key-value pairs are concatenated according to a preset field order to obtain a structured sequence. For unstructured text data, word segmentation based on a Chinese word segmentation algorithm is used. A custom dictionary for university finance and procurement is loaded, which includes exclusive vocabulary such as university budget subject names, procurement item names, internal department names, and financial professional terms. After word segmentation, stop words are removed. Stop words include auxiliary words, prepositions, conjunctions, and other words without actual semantic meaning, resulting in a text word segmentation sequence. The structured sequence and the text word segmentation sequence are concatenated end to end. An identifier is added at the beginning of the concatenated sequence, and a [SEP] identifier is added at the boundary between the structured sequence and the text word segmentation sequence. A [SEP] identifier is also added at the end of the concatenated sequence. At the same time, a first-type segment identifier is assigned to the structured sequence with a value of 0, and a second-type segment identifier is assigned to the text word segmentation sequence with a value of 1, resulting in the input sequence for the pre-trained language model. The pre-trained language model adopts a Chinese pre-trained language model based on the BERT architecture. The model has 12 hidden layers with a 768-dimensional hidden layer and 12 attention heads. The input sequence carrying segment identifiers is input into the pre-trained language model. After multi-layer Transformer encoding processing, the output vector corresponding to the identifier in the last hidden layer of the pre-trained language model is extracted as the execution feature vector, which has a 768-dimensional dimension.

[0023] A cross-modal attention matching network based on a sliding time window is constructed. The budget feature matrix is ​​used as the key and value vectors of the cross-modal attention matching network, and the execution feature vector is used as the query vector. The cross-modal attention matching network is built based on a scaled dot product attention mechanism. The network input includes the query vector, the key vector matrix, and the value vector matrix. The budget feature matrix is ​​linearly mapped by two independent fully connected layers. The first fully connected layer maps the dimension of the budget feature matrix from N×112 to N×768, obtaining the key vector matrix. The second fully connected layer maps the dimension of the budget feature matrix from N×112 to N×768, obtaining the value vector matrix. The dimensions of both the key vector matrix and the value vector matrix match the dimension of the execution feature vector. The preset duration of the sliding time window is 30 calendar days. Starting from the current system time, the window advances backward by 30 calendar days, collecting all procurement execution flows within that time range to form a procurement execution flow set for the current time window. Each flow in the set is sorted in descending order of its generation timestamp, with newer flows appearing earlier. The execution feature vectors corresponding to each flow in the sorted set are stacked to generate a query vector matrix. The query vector matrix has a dimension of M×768, where M is the total number of procurement execution flows within the current time window. The key vector matrix, value vector matrix, and query vector matrix are input into a cross-modal attention matching network. The key and value vector matrices only need to be updated synchronously based on budget item data updates within each computation cycle, without needing to be recalculated with each procurement execution flow input.

[0024] In the forward propagation process of the cross-modal attention matching network, a procurement stage masking mechanism is introduced. Based on the procurement lifecycle stage of the procurement execution flow, the weight coefficients of different feature dimensions in the cross-modal attention matching network are dynamically adjusted, outputting the matching probability distribution between the execution feature vector and the budget feature matrix. The procurement lifecycle stages include the demand submission stage, the contract signing stage, and the payment stage. Each procurement execution flow carries a corresponding procurement lifecycle stage label, which is synchronously assigned by the procurement business system when the flow is generated. The flow corresponding to the demand submission stage includes procurement demand application forms and procurement project initiation application forms; the flow corresponding to the contract signing stage includes procurement contract filing data and bid-winning notice data; and the flow corresponding to the payment stage includes procurement expense reports, payment application forms, and bank receipts. A pre-constructed stage embedding vector lookup table is built, containing a number of stage embedding vectors equal to the number of procurement lifecycle stage labels. Each stage label corresponds to a unique stage embedding vector, and the stage embedding vector has a dimension of 768, consistent with the dimension of the execution feature vector. Based on the stage label of the current procurement execution flow, the corresponding stage vector is retrieved from the stage embedding vector lookup table. The stage vector is then mapped using a linear fully connected layer. The input dimension of the linear fully connected layer is 768-dimensional, and the output dimension is M×N-dimensional, where M is the number of rows in the query vector matrix and N is the number of rows in the key vector matrix. This generates a mask tensor with the same dimension as the attention weight matrix of the cross-modal attention matching network. During the forward propagation of the cross-modal attention matching network, the dot product of the query vector and the key vector is calculated to obtain the original attention score matrix. The mask tensor is then added element-wise to the original attention score matrix to obtain the masked attention score matrix. The masked attention score matrix is ​​then scaled by a scaling factor equal to the reciprocal of the square root of the key vector dimension. The scaled attention score matrix is ​​then normalized using softmax to obtain weight coefficients for different feature dimensions. The value vector matrix is ​​then weighted and summed using these weight coefficients to obtain a fused feature vector that integrates budget and execution features. This fused feature vector is then input into a fully connected layer and a softmax normalization function to output the matching probability distribution of the current execution feature vector corresponding to each budget item. Each probability value in the matching probability distribution corresponds to a budget item, and the probability values ​​range from 0 to 1. The sum of all probability values ​​is 1.

[0025] Based on the matching probability distribution and a preset probability threshold, a mapping is established between procurement execution flows and corresponding budget items. The preset probability threshold is 0.85. All probability values ​​in the matching probability distribution are iterated through, and target probability values ​​greater than the preset threshold are selected. If the number of selected target probability values ​​is 1, the budget item corresponding to that target probability value is designated as the target budget item, and a mapping is established between the current procurement execution flow and the target budget item. In the database of the integrated financial management platform, a foreign key association is established between the flow number of the procurement execution flow and the item code of the target budget item, and the executed amount and available balance of the target budget item are updated synchronously. If the number of selected target probability values ​​is greater than 1, all budget items corresponding to the target probability values ​​are pushed to the operation terminal of the financial auditors in descending order of probability value for manual confirmation. The mapping is established based on the manual confirmation result. If no target probability value greater than the preset probability threshold is selected, the current procurement execution flow is marked as pending matching and pushed to the operation terminal of the budget management department for manual matching by budget management personnel.

[0026] Table 1. Example of structured form data fields and serialization concatenation for procurement execution flow.

[0027] This table illustrates the specific implementation of key-value pair serialization and concatenation of structured form data in this embodiment. It clarifies the concatenation rules and delimiter usage for different types of fields. By concatenating key-value pairs in a fixed order, it ensures that the semantics of the fields and the relationships between fields in the structured form data can be fully preserved. This provides a standardized structured sequence input for the encoding processing of pre-trained language models, avoiding the loss of semantic information due to disordered field order or inconsistent format.

[0028] In this embodiment, by converting the structured attribute data of university budget items into a budget feature matrix and converting the structured and unstructured data of procurement execution flow into a unified-dimensional execution feature vector, the mapping of structured budget data and unstructured execution data in the same feature space is realized, eliminating the semantic gap between the two types of data. Through a cross-modal attention matching network, deep matching of budget features and execution features is achieved. Combined with a procurement stage masking mechanism to dynamically adjust feature weights, the output matching probability distribution can accurately reflect the degree of correlation between execution flow and budget items. The finally established correlation mapping can effectively reduce the probability of mismatch and correlation failure, meeting the business needs of university procurement budget and execution linkage management.

[0029] In an alternative embodiment, refer to Figure 2The system extracts the subject codes, superior subject codes, and subject hierarchy attributes of all valid budget subjects from the budget management system. University budget subjects adopt a hierarchical management model, divided into first-level, second-level, third-level, and fourth-level budget subjects. Some universities can expand to fifth-level budget subjects according to management needs. Each lower-level budget subject uniquely corresponds to one superior budget subject, and a superior budget subject can correspond to multiple lower-level budget subjects, forming a tree-like hierarchical relationship. The subject codes use a segmented coding rule: the first-level subject code is 4 digits, the second-level subject code is 2 digits longer than the first-level code, the third-level subject code is 2 digits longer than the second-level code, and so on. Through the segmented rule of the subject codes and the superior subject code field, the hierarchical relationship of each budget subject node is analyzed. Using all first-level budget items as the root node of the tree topology, the second-level budget items corresponding to each root node are the child nodes of the root node, the third-level budget items corresponding to the second-level budget items are the child nodes of the second-level budget items, and so on, a directed acyclic budget item tree topology is constructed. Each node in the tree topology corresponds to a unique budget item, and the directed edges between nodes correspond to the hierarchical relationship between upper and lower level items. The direction of the directed edges is from the upper level item node to the lower level item node.

[0030] The budget item tree topology is traversed to obtain the item category code, management department code, and funding source code of each budget item node as structured attribute data. A depth-first traversal algorithm is used, starting from the root node of the tree topology and sequentially traversing all child nodes of each root node until all leaf nodes are traversed. A unique node index is assigned to each node during the traversal, and this node index is used to construct the adjacency matrix of the graph convolutional network. For each traversed budget item node, the corresponding item category code, management department code, and funding source code are extracted. For non-leaf nodes, their management department code and funding source code inherit from the union of all their child nodes. If multiple child nodes of a parent item node belong to different management departments, the management department code of the parent node is the set of the management department codes of all its child nodes, ensuring that the structured attribute data of each node in the tree topology can fully reflect the attribute characteristics of the item and its subordinate items.

[0031] The subject category code, the responsible management department code, and the funding source code are concatenated to obtain the initial feature vector for each budget subject node. The subject category code is converted into a numerical variable using label encoding, and then mapped to a 64-dimensional dense vector through an embedding layer. The responsible management department code is converted into a 32-dimensional vector using multi-hot encoding, where each dimension corresponds to a secondary management department within the school; if the subject node belongs to the corresponding department, the dimension has a value of 1, otherwise it is 0. The funding source code is also converted into a 16-dimensional vector using multi-hot encoding, where each dimension corresponds to a funding source type; if the subject node contains the corresponding funding source, the dimension has a value of 1, otherwise it is 0. The dense vector corresponding to the subject category code, the multi-hot vector corresponding to the responsible management department code, and the multi-hot vector corresponding to the funding source code are concatenated in a preset order to obtain the initial feature vector for each budget subject node, with an initial feature vector dimension of 112.

[0032] A graph convolutional network is used to perform feature aggregation calculations on the budget item tree topology graph and the initial feature vectors. The updated hidden state of each budget item node is used as the row vector in the budget feature matrix. An adjacency matrix A is constructed corresponding to the budget item tree topology graph, with dimensions [missing information]. Where N is the total number of nodes in the tree topology graph, and if there is a directed edge between node i and node j, then in the adjacency matrix... ,otherwise Simultaneously, an identity matrix I is added to the adjacency matrix to obtain a self-loop adjacency matrix. This ensures that each node retains its initial features during the feature aggregation process.

[0033] For any central node v in the budget subject tree topology graph, obtain the set of neighboring nodes of central node v within a preset hop count range. The preset hop count is 2 hops, which includes the direct child nodes, direct parent nodes, and other child nodes of the parent node of the central node, ensuring that the feature information of adjacent levels in the subject hierarchy can be aggregated. Extract the initial feature vector corresponding to each neighbor node in the set of neighboring nodes, and multiply each initial feature vector by the corresponding edge weight coefficient. The edge weight coefficient is calculated by decaying based on the shortest path length between two connected nodes in the budget subject tree topology graph. The calculation formula is as follows: ; in, The edge weight coefficient between node u and node v. Let be the shortest path length between node u and node v in the tree topology graph. The shortest path length is calculated using the breadth-first search algorithm. The path length is a positive integer. When node u and node v are the same node, the shortest path length is 0 and the edge weight coefficient is 1 to ensure that the weight of the node's own features is the highest. As the path length increases, the edge weight coefficient decreases inversely to reduce the influence of distant nodes on the feature aggregation of the central node.

[0034] Summing the multiplied initial feature vectors yields the neighbor aggregation feature of the central node, calculated using the following formula: ; in, The neighbor aggregation features of the central node v The set of neighboring nodes within a preset hop count range of the central node v. Let be the initial feature vector corresponding to the neighbor node u.

[0035] The initial feature vector of the center node is concatenated with the aggregated features of its neighbors. The concatenated result is then input into an activation function to obtain the updated hidden state of the center node. The calculation formula is as follows: ; in, This represents the hidden state of the node after the central node v has been updated. It is the ReLU activation function. For a trainable weight matrix, is a trainable bias vector. This is a vector concatenation operation. Let v be the initial feature vector of the central node v.

[0036] The graph convolutional network is set to have two layers. After two layers of graph convolution computation, the final hidden state of each budget item node is obtained. The final hidden states of all nodes are arranged in order of node index to generate the budget feature matrix. The dimensions of the budget feature matrix are... The dimension is then transformed into 768 dimensions that match the dimension of the execution feature vector through a linear mapping layer for cross-modal attention matching computation.

[0037] Table 2 Examples of Node Attributes and Initial Feature Vectors in the Tree Topology Diagram of Budget Items

[0038] This table shows the correspondence between the structured attribute data of nodes at different levels and the initial feature vectors in the tree topology diagram of budget subjects. It verifies that the attribute data of nodes at different levels can be effectively encoded into initial feature vectors of fixed dimensions. The attribute data of the root node and non-leaf nodes inherit the attribute union of their subordinate child nodes, ensuring that the initial features of each node in the tree topology diagram can fully reflect the full attribute information of the corresponding subject and its subordinate subjects, providing a standardized input for feature aggregation of the subsequent graph convolutional network.

[0039] In this embodiment, by constructing a tree-like topology graph of budget items, the hierarchical relationship between budget items is fully preserved. By using a graph convolutional network to aggregate the features of budget item nodes in multiple layers, the generated budget feature matrix not only contains the attribute features of a single budget item, but also integrates the hierarchical relationship between items and the association features of adjacent items. This solves the problem that a single item feature cannot reflect the hierarchical relationship, improves the expressive power of budget features, and provides a richer feature foundation for subsequent cross-modal matching.

[0040] In an alternative embodiment, refer to Figure 3 The system predefines the field priority order of the structured form data, ranking them from highest to lowest based on their relevance to the budget item. The order is: Procurement Project Name, Procurement Item Code, Applying Department, Fund Purpose, Procurement Method, Procurement Amount, Contract Number, Invoice Number, Applicant, and Application Time. Following the predefined field priority order, the system sequentially reads the field name and corresponding value for each field. Fields with empty values ​​are skipped to avoid meaningless null values ​​affecting the semantic expression of the sequence. For each non-empty field, the field name and value are concatenated using a colon ":", and fields are separated by semicolons ";" to ensure the pre-trained language model can recognize the boundaries between different fields, resulting in a structured sequence.

[0041] Unstructured text data is segmented to obtain a text segmentation sequence. The unstructured text data includes multiple segments such as procurement requirements, contract text, invoice remarks, and reimbursement instructions. First, these segments are concatenated in paragraph order to form a complete unstructured text string. A segmenter based on a pre-trained language model is then used to segment the unstructured text string. The segmenter is BERT-Chinese, which supports segmentation of Chinese words, English words, numbers, and punctuation marks. During segmentation, rare characters and special symbols are automatically handled, and unrecognizable characters are converted to [UNK] identifiers. After word segmentation, a text segmentation sequence is obtained. Each element in the text segmentation sequence is a segmentation token. The number of tokens does not exceed the maximum input length limit of the pre-trained language model. The maximum input length is set to 512 tokens. If the number of tokens after segmentation exceeds 512, the text segmentation sequence is truncated, retaining the first 400 tokens and discarding the remaining tokens to ensure that the total length of the concatenated sequence does not exceed the limit of 512 tokens.

[0042] The structured sequence and the text segmentation sequence are concatenated, with the structured sequence assigned a first-class segment identifier and the text segmentation sequence assigned a second-class segment identifier. The structured sequence is segmented using the same tokenizer to obtain a structured segmentation sequence, with no more than 112 tokens. The total number of tokens after concatenating the structured and text segmentation sequences is ensured not to exceed 512. An identifier is added at the beginning of the structured segmentation sequence, and a [SEP] identifier is added at the end of the structured segmentation sequence to form a complete structured segmentation subsequence. Similarly, a [SEP] identifier is added at the end of the text segmentation sequence to form a complete text segmentation subsequence. The structured and text segmentation subsequences are then concatenated end-to-end to obtain the complete input segmentation sequence. Each token in the input segmented sequence is assigned a corresponding segment identifier. In the structured segmented subsequence, all tokens, including the identifier and the first [SEP] identifier, are assigned a first-class segment identifier with a value of 0. In the text segmented subsequence, all tokens, including the second [SEP] identifier, are assigned a second-class segment identifier with a value of 1. Simultaneously, each token in the input segmented sequence is assigned a position identifier, which increments sequentially from 0 according to the token's position in the sequence. The pre-trained language model uses these position identifiers to identify the token's position within the sequence.

[0043] The concatenated sequence carrying both first-class and second-class segment identifiers is input into the pre-trained language model. The feature vector corresponding to the first-class segment identifier in the last hidden layer of the pre-trained language model is extracted as the execution feature vector. The Transformer encoder layer of the pre-trained language model encodes the input sequence through a multi-head attention mechanism, capturing the semantic relationships between different tokens in the sequence. The structured sequence corresponding to the first-class segment identifier contains the core structured information of the procurement execution flow and has a higher matching correlation with the structured attribute data of the budget items. Therefore, extracting the output vector of the identifier corresponding to the first-class segment identifier as the execution feature vector can more accurately express the core matching features of the procurement execution flow. The output dimension of the last hidden layer of the pre-trained language model is 768-dimensional; therefore, the extracted execution feature vector is a dense vector of 768 dimensions, which can completely preserve the semantic features of both the structured and unstructured data of the procurement execution flow.

[0044] Before extracting the feature vectors corresponding to the first-class segment identifiers in the last hidden layer of the pre-trained language model, a historical procurement execution flow sample set and corresponding historical budget item sample labels are obtained. The historical procurement execution flow sample set comes from the procurement execution flow data of the past three fiscal years in the university's financial system that has been matched. Each sample contains structured form data, unstructured text data, and the corresponding correctly matched budget item code as a sample label. The historical procurement execution flow sample set is cleaned to remove invalid samples with mismatches and missing data, resulting in 100,000 valid samples. Among them, the number of samples corresponding to each budget item label is no less than 10, ensuring the balance of sample distribution.

[0045] The historical procurement execution flow sample set is input into the pre-trained language model to extract sample feature vectors. Following the aforementioned sequence concatenation and segment identifier allocation method, the structured form data and unstructured text data of each sample are processed to generate an input sequence. This input sequence is then input into the pre-trained language model to extract the corresponding sample feature vector for each sample. The sample feature vector is a 768-dimensional vector, consistent with the dimension of the execution feature vector.

[0046] Using sample feature vectors as anchors, positive sample feature vectors belonging to the same historical budget item label are considered positive examples, and negative sample feature vectors belonging to different historical budget item labels are considered negative examples. A contrastive learning loss function is calculated. In each training batch, 32 samples are randomly selected as a batch, and each sample in the batch serves as an anchor sample. Within the same batch, other samples with the same budget item label as the anchor sample are considered positive samples, and samples with different budget item labels are considered negative samples. Each anchor sample corresponds to at least one positive sample and at least 30 negative samples. The contrastive learning loss function uses the InfoNCE loss function scaled by the temperature coefficient, and the calculation formula is as follows: ; in, To compare the learning loss function values, This represents the total number of samples in the training batch. Let i be the feature vector of the sample corresponding to anchor point sample i. Let i be the positive sample feature vector corresponding to anchor sample i. Let j be the feature vector corresponding to the j-th sample in the batch, including both positive and negative samples. This is the cosine similarity calculation function. The temperature coefficient, with a value of 0.07, is used to scale the range of cosine similarity values ​​and control the model's ability to distinguish between positive and negative samples.

[0047] The formula for calculating cosine similarity is: ; in, and Let's consider two feature vectors whose similarity is to be calculated. For the dot product operation of vectors, Let L2 be the norm of the vector.

[0048] The model parameters of the pre-trained language model are updated using backpropagation based on the gradient of the contrastive learning loss function. The AdamW optimizer is used for parameter updates, with the learner rate set to a specific value. The weight decay coefficient is set to The training epochs were set to 3, and the batch size to 32. In each training step, the contrastive learning loss function was calculated via forward propagation, and then the gradient of the loss function with respect to the parameters of each layer of the model was calculated via backpropagation. The model parameters were then updated according to the optimizer's update rules. An early stopping mechanism was used during training: when the loss function value on the validation set no longer decreased for 100 consecutive training steps, training was stopped, the optimal model parameters were saved, and the fine-tuning of the pre-trained language model was completed.

[0049] Table 3. Examples of input sequence structure and segment identifier allocation for pre-trained language models

[0050] This table illustrates the complete structure of the input sequence for the pre-trained language model, clarifying the allocation rules for token content, segment identifiers, and position identifiers. By assigning different segment identifiers to structured sequences and text segmentation sequences, the pre-trained language model can effectively distinguish between two different types of data and capture the semantic features of each type. At the same time, the position identifiers preserve the positional information of the tokens in the sequence, ensuring that the model can accurately understand the contextual semantics of the sequence and providing a standardized input structure for the precise extraction of feature vectors.

[0051] In this embodiment, by performing key-value pair serialization and concatenation on structured form data and standardized word segmentation on unstructured text data, and assigning corresponding segment identifiers to different types of sequences, the pre-trained language model can simultaneously encode the semantic features of both structured and unstructured data, solving the problem that the two types of data cannot be simultaneously input into the model for feature extraction. At the same time, the pre-trained language model is fine-tuned for the field of university financial procurement through comparative learning, enabling the model to better capture domain-specific semantic features, improving the domain adaptability and discriminativeness of the executed feature vectors, and providing accurate feature input for subsequent cross-modal matching.

[0052] In an alternative embodiment, refer to Figures 4 to 6 This process retrieves the current set of procurement execution logs and sorts them in descending order based on their generation timestamps. The current time is the server system's real-time clock. The procurement execution log set comprises all generated but not yet budgeted procurement execution logs from the university's procurement business system. Each log carries a unique generation timestamp, accurate to the second, identifying its creation time. All logs in the set are sorted in descending order of their generation timestamp values; larger timestamps indicate later creation and are placed earlier in the list, ensuring that the most recently generated logs are processed first.

[0053] Using the current moment as the end point of the time window and a preset time length as the time window span, a subset of target transaction flows within the time window is extracted from the sorted procurement execution flow set. The initial value of the preset time length is 30 calendar days, the preset baseline number of transactions is 100, and the fixed interval is 80 to 120 transactions. First, the number of historical transactions within the three adjacent historical time windows before the current moment is counted, and the average of the historical transaction flows within the three historical time windows is calculated as the statistical value of the historical transaction flow count. The difference between the statistical value of the historical transaction flow count and the preset baseline number of transactions is calculated. If the difference is greater than zero, it means that the number of transactions within the historical time window exceeds the baseline number, and the data processing pressure is high. In this case, the preset time length is shortened proportionally to the difference. If the difference is less than zero, it means that the number of transactions within the historical time window is lower than the baseline number, and the data processing pressure is low. In this case, the preset time length is extended proportionally to the difference, resulting in the adjusted actual time window span. The formula for calculating the actual time window span is: ; in, This refers to the adjusted actual time window span, in calendar days. The initial value is the preset time length. To preset the baseline flow rate, This represents the statistical value of historical transaction volume within adjacent historical time windows.

[0054] Based on the actual time window span, and using the current moment as the endpoint, the system traces back to the corresponding natural days within the actual time window span. From the sorted procurement execution log set, logs whose generation timestamps fall within the traceback time range are selected to form the target log subset. By dynamically adjusting the time window span, the number of log entries in the target log subset is kept within a fixed range of 80 to 120, avoiding data processing delays caused by an excessive number of log entries, while also avoiding statistical biases caused by an insufficient number of log entries.

[0055] The execution feature vectors corresponding to the target pipeline subset are used to form a query vector matrix. The budget feature matrix is ​​divided into a key vector matrix and a value vector matrix that match the dimension of the query vector matrix. The key vector matrix, value vector matrix, and query vector matrix are then input into a cross-modal attention matching network. The target pipeline subset contains M procurement execution pipelines, each corresponding to a 768-dimensional execution feature vector. The M execution feature vectors are stacked according to the pipeline arrangement to generate a query vector matrix Q, which has a dimension of M×768. The budget feature matrix has a dimension of N×128, where N is the total number of valid budget items. The budget feature matrix is ​​linearly mapped through two independent fully connected layers. The first fully connected layer has a weight matrix dimension of 128×768, mapping the budget feature matrix to an N×768-dimensional key vector matrix K. The second fully connected layer has a weight matrix dimension of 128×768, mapping the budget feature matrix to an N×768-dimensional value vector matrix V. The key vector matrix K, value vector matrix V, and query vector matrix Q all have a feature dimension of 768, which satisfies the input dimension matching requirement of the scaled dot product attention mechanism. The three matrices are then input into the cross-modal attention matching network for subsequent attention calculation.

[0056] The procurement execution flow is tagged with the current procurement lifecycle stage. These stages include the demand submission stage, contract signing stage, and payment stage, each corresponding to the entire procurement process. The core matching features differ for each stage: the demand submission stage's core matching features are the name and purpose of the procurement project, with the highest correlation to the budget item name; the contract signing stage's core matching features are the procurement item and contract amount, with the highest correlation to the budget item category and budget amount; and the payment stage's core matching features are the payment amount and purpose of funds, with the highest correlation to the budget item's funding source and available balance. A stage embedding vector lookup table, equal in number to the procurement lifecycle stage tags, is constructed. This table contains three stage embedding vectors, each corresponding to one of the three procurement lifecycle stage tags. Each stage embedding vector has 768 dimensions. These vectors are learned and updated end-to-end during model training, accurately representing the feature preferences of different procurement stages. Based on the procurement lifecycle stage label, the corresponding stage vector is retrieved from the stage embedding vector lookup table. Each stage vector has the same dimension as the execution feature vector of the corresponding pipeline.

[0057] A linear fully connected mapping is applied to the stage vectors to generate a mask tensor with the same dimension as the attention weight matrix of the cross-modal attention matching network. This mask tensor is then superimposed onto the attention score matrix of the cross-modal attention matching network. The execution feature vectors are divided into three subspaces: name-dimensional, amount-dimensional, and time-dimensional, with a total dimension of 768. The name-dimensional feature subspace corresponds to the first 256 dimensions of the execution feature vectors, representing textual semantic features such as the procurement project name, procurement item, and purpose. The amount-dimensional feature subspace corresponds to the middle 256 dimensions, representing numerical features such as the procurement amount and budget. The time-dimensional feature subspace corresponds to the last 256 dimensions, representing time-related features such as the procurement time and budget year. These three feature subspaces have equal and non-overlapping dimensions and correspond to the three core matching features of the procurement execution flow.

[0058] The stage vector is processed by a linear fully connected mapping. The input dimension of the linear fully connected layer is 768-dimensional, and the output dimension is M×N-dimensional, where M is the number of flow items in the target flow subset, N is the total number of budget items, and the output mask tensor has a dimension of M×N-dimensional, which is completely consistent with the dimension of the attention score matrix of the cross-modal attention matching network. During the linear fully connected mapping process, the mapping values ​​corresponding to different feature subspaces are dynamically set according to the different labels of the procurement lifecycle stage. When the procurement lifecycle stage label is the demand declaration stage, the mapping values ​​corresponding to the name dimension feature subspace in the stage vector are set to positive numbers and the mapping values ​​corresponding to the amount dimension feature subspace are set to negative numbers through the linear fully connected mapping. This increases the feature weight of the name dimension and suppresses the feature weight of the amount dimension during the attention calculation process. When the procurement lifecycle stage label is the payment stage, the mapping values ​​corresponding to the amount dimension feature subspace in the stage vector are set to positive numbers and the mapping values ​​corresponding to the name dimension feature subspace are set to negative numbers through the linear fully connected mapping. This increases the feature weight of the amount dimension and suppresses the feature weight of the name dimension during the attention calculation process. When the procurement lifecycle stage label is the contract signing stage, the mapping values ​​corresponding to both the name and amount dimension feature subspaces in the stage vector are set to positive numbers, while the mapping value corresponding to the time dimension feature subspace is set to zero, balancing the feature weights of the name and amount dimensions. A mask tensor is generated based on the mapped values ​​after setting the values. Each element in the mask tensor corresponds to the weight adjustment coefficient at the corresponding position in the attention score matrix.

[0059] The weight coefficients of different feature dimensions in the cross-modal attention matching network are dynamically adjusted, and the matching probability distribution between the execution feature vector and the budget feature matrix is ​​output. The dot product of the query vector and the key vector is calculated to obtain the original attention score matrix, and the calculation formula is as follows: ; in, The original attention score matrix has the following dimensions: Dimension, Q is the query vector matrix with dimension . Dimension K is the key vector matrix with dimension 1. dimension, Let be the transpose of the key vector matrix, with dimension . dimension.

[0060] The original attention score matrix is ​​added element-wise to the mask tensor to obtain the mask attention score matrix. The calculation formula is as follows: ; in, The mask attention score matrix has dimensions of . Dimension M is the mask tensor with dimension 1. The + sign indicates element-wise addition.

[0061] The mask attention score matrix is ​​scaled using a scaling factor equal to the reciprocal of the square root of the key vector dimension. This prevents the dot product from becoming too large, which could cause the gradient of the softmax function to vanish. The calculation formula is as follows: ; in, This is the scaled attention score matrix. This represents the feature dimension of the key vector, with a value of 768. It is the square root of the dimension of the key vector.

[0062] Normalized exponentiation is performed on each row of the scaled mask attention score matrix to generate weight coefficients. The calculation formula is as follows: ; Where A is the attention weight matrix, with dimension 1. The softmax function performs a normalized exponential operation row by row, where the sum of all elements in each row is 1, and each element corresponds to the weight coefficient of a budget item.

[0063] The fused feature vector is obtained by weighting and summing the value vectors using weighting coefficients. The calculation formula is as follows: ; Where F is the fusion feature matrix composed of fused feature vectors, with dimension . Dimension V is a value vector matrix with dimension 1. Each row vector in the dimension is a fused feature vector of the corresponding procurement execution flow, which integrates budget features and execution features, and has undergone weight adjustment in the procurement stage.

[0064] The fused feature vectors are input into a fully connected layer and a normalized exponential function, and the output feature vectors correspond to the matching probability distributions of each budget item. The fully connected layer has an input dimension of 768 and an output dimension of N, where N is the total number of budget items. The fused feature matrix F is input into the fully connected layer to obtain a feature vector of dimension F. The output matrix is ​​dimensional. A softmax normalization exponent operation is performed on each row of the output matrix to obtain the matching probability distribution of each procurement execution flow corresponding to each budget item. Each probability value in the matching probability distribution corresponds to a budget item. The probability value ranges from 0 to 1, and the sum of all probability values ​​in each row is 1.

[0065] Based on the matching probability distribution and preset probability thresholds, a mapping between procurement execution flow and corresponding budget items is established. The preset probability thresholds are dynamically adjusted according to different stages of the procurement lifecycle: 0.9 for the demand submission stage, 0.85 for the contract signing stage, and 0.8 for the payment stage. All probability values ​​in the matching probability distribution are iterated through, and target probability values ​​greater than the corresponding stage's preset probability threshold are selected. If the number of selected target probability values ​​is 1, the budget item corresponding to that target probability value is designated as the target budget item, and a mapping between the current procurement execution flow and the target budget item is established. If the number of selected target probability values ​​is greater than 1, all budget items corresponding to the target probability values ​​are pushed to the financial auditor's terminal in descending order of probability value for manual confirmation. If no target probability value greater than the preset probability threshold is selected, the current procurement execution flow is marked as pending matching and pushed to the budget management department for manual processing.

[0066] Table 4. Mask tensor setting rules and feature weight adjustment examples for different procurement lifecycle stages.

[0067] This table shows the mask tensor setting rules corresponding to different procurement lifecycle stages, and clarifies the weight adjustment methods for different feature subspaces at different stages. By setting positive and negative values ​​for the mapping values ​​of different feature subspaces, dynamic adjustment of the weights of different feature dimensions during attention calculation is achieved, enabling the cross-modal attention matching network to adapt to the business needs of different procurement stages and ensuring that the matching results conform to the full-process management logic of university procurement business.

[0068] In this embodiment, by dynamically adjusting the span of the sliding time window, the number of input streams is kept stable, balancing the real-time performance and stability of data processing and reducing the latency of batch data processing. By introducing a procurement stage masking mechanism, the weight coefficients of different feature dimensions are dynamically adjusted according to different stages of the procurement lifecycle, enabling the cross-modal attention matching network to adapt to the business feature requirements of different procurement stages, improving the accuracy and business adaptability of the matching results. The final output matching probability distribution can accurately reflect the correlation between the procurement execution stream and the budget item, providing a reliable basis for establishing the correlation mapping and realizing the full-process linkage management of procurement budget and execution.

Claims

1. A method for linking university procurement budget and execution management, characterized in that, include: Extract structured attribute data of university budget items and generate a budget feature matrix; Collect unstructured text data and structured form data of the procurement execution flow, input the unstructured text data and the structured form data into a pre-trained language model, and extract execution feature vectors; A cross-modal attention matching network based on a sliding time window is constructed, with the budget feature matrix used as the key vector and value vector of the cross-modal attention matching network, and the execution feature vector used as the query vector of the cross-modal attention matching network. During the forward propagation of the cross-modal attention matching network, a procurement stage masking mechanism is introduced. Based on the procurement lifecycle stage of the procurement execution flow, the weight coefficients of different feature dimensions in the cross-modal attention matching network are dynamically adjusted, and the matching probability distribution of the execution feature vector and the budget feature matrix is ​​output. Based on the matching probability distribution and the preset probability threshold, an association mapping is established between the procurement execution flow and the corresponding budget item.

2. The method for linking university procurement budget and execution management according to claim 1, characterized in that, The step of extracting structured attribute data of university budget items and generating a budget feature matrix includes: parsing the hierarchical relationship between university budget items and constructing a tree topology diagram of budget items; Traverse the tree topology of the budget items and obtain the item category code, the management department code, and the funding source code of each budget item node as the structured attribute data; The subject category code, the department code, and the funding source code are concatenated to obtain the initial feature vector of each budget subject node; The graph convolutional network is used to perform feature aggregation calculation on the tree topology graph of the budget items and the initial feature vector, and the updated hidden state of each budget item node is used as the row vector in the budget feature matrix.

3. The method for linking university procurement budget and execution management according to claim 1, characterized in that, The step of inputting the unstructured text data and the structured form data into a pre-trained language model and extracting execution feature vectors includes: performing key-value pair serialization concatenation on the field names and field values ​​in the structured form data to obtain a structured sequence; The unstructured text data is segmented into words to obtain a text segmentation sequence; The structured sequence is concatenated with the text segmentation sequence, and a first type of segment identifier is assigned to the structured sequence and a second type of segment identifier is assigned to the text segmentation sequence in the concatenated sequence. The concatenated sequence carrying the first segment identifier and the second segment identifier is input into the pre-trained language model, and the feature vector corresponding to the first segment identifier in the last hidden layer of the pre-trained language model is extracted as the execution feature vector.

4. The method for linking university procurement budget and execution management according to claim 1, characterized in that, The construction of the cross-modal attention matching network based on a sliding time window includes: obtaining the current procurement execution flow set and sorting it in descending order according to the generation timestamp of each flow in the procurement execution flow set; Using the current moment as the end point of the time window and a preset time length as the span of the time window, a subset of the target flow within the time window is extracted from the sorted procurement execution flow set. The execution feature vectors corresponding to the target pipeline subset are used to form a query vector matrix; The budget feature matrix is ​​divided into a key vector matrix and a value vector matrix that match the dimension of the query vector matrix. The key vector matrix, the value vector matrix, and the query vector matrix are then input into the cross-modal attention matching network.

5. The method for linking university procurement budget and execution management according to claim 1, characterized in that, The introduced procurement stage masking mechanism includes: obtaining the procurement lifecycle stage label of the current procurement execution flow, wherein the procurement lifecycle stage label includes the demand submission stage, contract signing stage, and fund payment stage; Construct a stage embedding vector lookup table with the same number of procurement lifecycle stage tags, and retrieve the corresponding stage vector from the stage embedding vector lookup table based on the procurement lifecycle stage tags; A linear fully connected mapping is performed on the stage vector to generate a mask tensor with the same dimension as the attention weight matrix of the cross-modal attention matching network. The mask tensor is then superimposed on the attention score matrix of the cross-modal attention matching network.

6. The method for linking university procurement budget and execution management according to claim 5, characterized in that, The step of dynamically adjusting the weight coefficients of different feature dimensions in the cross-modal attention matching network and outputting the matching probability distribution between the execution feature vector and the budget feature matrix includes: calculating the dot product of the query vector and the key vector to obtain the original attention score matrix; The original attention score matrix is ​​added element-wise to the mask tensor to obtain the mask attention score matrix; Perform a normalized exponential operation on each row of the mask attention score matrix to generate the weight coefficients; The value vector is weighted and summed using the weighting coefficients to obtain a fused feature vector. The fused feature vector is input into a fully connected layer and a normalized exponential function, and the matching probability distribution of the execution feature vector corresponding to each budget item is output.

7. The method for linking university procurement budget and execution management according to claim 2, characterized in that, The step of using a graph convolutional network to perform feature aggregation calculation on the budget item tree topology graph and the initial feature vector includes: for any central node in the budget item tree topology graph, obtaining the set of neighboring nodes of the central node within a preset number of hops; Extract the initial feature vector corresponding to each neighbor node in the neighbor node set, and multiply each initial feature vector by the corresponding edge weight coefficient; The initial feature vectors after multiplication are summed to obtain the neighbor aggregation feature of the central node; The initial feature vector of the central node is concatenated with the aggregated features of its neighbors, and the concatenation result is input into an activation function to obtain the updated hidden state of the central node. The edge weight coefficient is calculated by decaying based on the shortest path length between two connected nodes in the budget subject tree topology graph.

8. The method for linking university procurement budget and execution management according to claim 3, characterized in that, Before extracting the feature vector corresponding to the first class segment identifier in the last hidden layer of the pre-trained language model, the method further includes: obtaining a historical procurement execution flow sample set and the corresponding historical budget item sample labels, inputting the historical procurement execution flow sample set into the pre-trained language model, and extracting sample feature vectors; Using the sample feature vector as the anchor point, positive sample feature vectors belonging to the same historical budget item label are taken as positive examples, and negative sample feature vectors belonging to different historical budget item labels are taken as negative examples, and the contrastive learning loss function is calculated. The model parameters of the pre-trained language model are updated based on the gradient backpropagation of the contrastive learning loss function, wherein the contrastive learning loss function adopts the InfoNCE loss function scaled by the temperature coefficient.

9. A method for linking university procurement budget and execution management according to claim 4, characterized in that, Using a preset time length as the time window span, extract a target flow subset within the time window from the sorted procurement execution flow set, including: counting the number of historical flows within adjacent historical time windows before the current moment; Calculate the difference between the historical transaction volume and the preset baseline transaction volume; If the difference is greater than zero, the preset time length is shortened by the proportion of the difference; if the difference is less than zero, the preset time length is extended by the proportion of the difference, thus obtaining the adjusted actual time window span. Based on the actual time window span, the target subset of the procurement execution flow is extracted from the sorted procurement execution flow set.

10. The method for linking university procurement budget and execution management according to claim 6, characterized in that, Performing a linear fully connected mapping on the stage vector to generate a mask tensor with the same dimension as the attention weight matrix of the cross-modal attention matching network includes: dividing the execution feature vector into a name dimension feature subspace, an amount dimension feature subspace, and a time dimension feature subspace; When the procurement lifecycle stage label is the demand declaration stage, the mapping value in the stage vector corresponding to the name dimension feature subspace is set to a positive number and the mapping value corresponding to the amount dimension feature subspace is set to a negative number through the linear fully connected mapping. When the procurement lifecycle stage label is the payment stage, the mapping value corresponding to the amount dimension feature subspace in the stage vector is set to a positive number through the linear fully connected mapping, and the mapping value corresponding to the name dimension feature subspace is set to a negative number. The mask tensor is generated based on the set mapping value.

Images