A technical requirement matching graph construction method based on a graph neural network

Abstract

The present application relates to the technical field of knowledge graph construction and information retrieval sorting, and particularly relates to a technical requirement matching graph construction method based on a graph neural network, comprising: reading technical requirement and technical supply literature, extracting term entities and generating candidate pairs according to string similarity and semantic similarity; constructing a candidate synonym cluster graph according to the candidate pairs and writing into a heterogeneous requirement matching graph, inputting a Bayesian graph neural network to obtain node embedding and node and edge uncertainty parameters; weighting and aggregating adjacent information based on edge uncertainty and updating within the cluster by applying consistency constraints; calculating a matching sorting score, writing requirement-supply matching edges according to a threshold and Top-K, recording confidence and explanation fields, and generating and storing a graph version. The present application improves requirement-supply matching accuracy under uncertainty constraints and reduces the mis-matching rate.

Description

Technical Field

[0001] This invention relates to the technical field of knowledge graph construction and information retrieval ranking, and in particular to a method for constructing a technology demand matching graph based on graph neural networks. Background Technology

[0002] In recent years, digital management and knowledge engineering for technological innovation activities have been continuously deepening, and technical information carriers such as R&D needs, technical solutions, and patent documents have shown a trend of large-scale, heterogeneous, and rapid evolution. Against this backdrop, technology demand matching methods based on natural language processing and knowledge graphs have gradually become the mainstream technical approach. On the one hand, unstructured technical texts are transformed into structured knowledge units through term extraction, entity disambiguation, and relation extraction. On the other hand, vectorized representation learning and similarity retrieval are used to recall and rank candidate documents from both the demand side and the supply side. In particular, the application of graph neural networks in knowledge graph representation learning enables node representations to integrate neighborhood structural information and express multiple types of relationships, thereby supporting tasks such as cross-document semantic association, technical element alignment, and matching recommendation.

[0003] However, existing technical requirement matching and graph construction solutions still have significant shortcomings in engineering implementation, particularly in terms of alignment noise and graph propagation error accumulation caused by synonyms / variants of terms: First, the same concept in technical documents often appears in the form of abbreviations, industry aliases, translation differences, parameter naming differences, etc., leading to missed recalls and misalignments during the cross-document alignment stage after term entity extraction, thus forming incorrect edges in the graph. This makes the subsequent message aggregation process based on graph neural networks susceptible to noise propagation, resulting in node embedding vector drift and unstable matching ranking; Second, traditional candidate screening based on similarity thresholds often... Similarity is treated as a deterministic quantity, lacking a mechanism to measure the reliability of candidate alignment and the credibility of edge relationships. This makes it difficult to distinguish between candidate edges with high similarity but low credibility and stable relationships with medium similarity but high credibility when multiple heterogeneous relationships are mixed, which can easily introduce uncontrollable errors in the graph construction stage. Thirdly, existing solutions often only output one-time matching results and lack a mechanism to write back the matching relationships as graph structural elements and manage them in a versioned manner. This makes it difficult to support subsequent retrieval traceability, incremental update reconciliation and model iteration evaluation, resulting in problems such as uncontrollable matching edge quality, untraceable graph evolution and large fluctuations in retrieval results in large-scale graph database scenarios. Summary of the Invention

[0004] In view of the problems of high alignment noise, accumulation of neighborhood propagation error, and lack of traceable versioning of matching results in existing technical requirements matching and map construction techniques under the condition of term synonyms / synonyms, this invention is proposed.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] In a first aspect, the present invention provides a method for constructing a technology demand matching graph based on a graph neural network, comprising: reading a set of technology demand literature texts and a set of technology supply literature texts, performing term entity extraction to obtain a term entity set, and generating a candidate pair set based on string similarity and semantic similarity;

[0007] Based on the candidate pair set, a candidate synonym cluster graph is constructed and written into the heterogeneous demand matching graph. The heterogeneous demand matching graph is then input into a Bayesian graph neural network to obtain node embedding vectors, node uncertainty measurement parameters, and edge uncertainty measurement parameters.

[0008] The adjacent edge information vectors are weighted and aggregated according to the edge uncertainty metric parameter, and a consistency constraint is applied within the cluster of the candidate synonym cluster graph to update the node embedding vector and the node uncertainty metric parameter.

[0009] Based on the updated node embedding vector, a matching ranking score is generated for demand nodes and supply nodes. The demand-supply matching edge is written into the heterogeneous demand matching graph according to the scoring threshold and Top-K rule of the matching ranking score. At the same time, a matching confidence field and an uncertainty measurement interpretation field are written into the demand-supply matching edge to generate an updated graph version and store it in the graph database.

[0010] In a second aspect, the present invention provides a computer device, comprising: one or more processors;

[0011] The memory stores operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the aforementioned graph neural network-based technical requirement matching graph construction method.

[0012] Thirdly, the present invention provides a computer-readable medium for storing software, the software including instructions executable by one or more computers, the instructions causing the one or more computers to perform operations, the operations including the process of the aforementioned graph neural network-based technical requirement matching graph construction method.

[0013] The beneficial effects of this invention are as follows: By reading the text sets of technical demand documents and technical supply documents and performing term entity extraction, this invention generates a candidate pair set based on string similarity and semantic similarity, enabling demand term entities and supply term entities to achieve high coverage and recall under multiple representation conditions, reducing missed recall and improving candidate quality; by constructing a candidate synonym cluster graph and writing it into a heterogeneous demand matching graph, and introducing the node embedding vector and node / edge uncertainty measurement parameters output by a Bayesian graph neural network, the graph representation simultaneously possesses semantic representation and reliability characterization capabilities; by weighting and aggregating the information vectors transmitted by adjacent edges according to the edge uncertainty measurement parameters and applying intra-cluster consistency constraints, the propagation of noise edges is suppressed and the representation drift caused by synonym variants is reduced, thereby improving matching stability; by generating a matching ranking score and writing it into the demand-supply matching edges and explanation fields according to the score threshold and Top-K rules, the matching relationship is versioned and solidified and traceable for retrieval, thereby improving the update consistency and retrieval reliability of the graph database. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0015] Figure 1 This is a flowchart illustrating the technical requirement matching graph construction method based on graph neural networks as shown in this invention. Detailed Implementation

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0017] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0018] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0019] According to an embodiment of the present invention, in combination Figure 1The flowchart shown illustrates a method for constructing a technical demand matching graph based on graph neural networks, which specifically includes the following steps:

[0020] S1. Read the text sets of technical requirement documents and technical supply documents, perform term entity extraction to obtain a term entity set, and generate a candidate pair set based on string similarity and semantic similarity. Note that the following points should be noted in this step:

[0021] S1.1 Perform encoding normalization and symbol standardization processing on the set of technical requirement documents and the set of technical supply documents to be read, and write document identifiers and fragment position identifiers to obtain a set of text fragments.

[0022] In a preferred embodiment, the technical requirement document text set consists of requirement-side source document entries, which include at least one of the following: enterprise technical requirement collection form, government / park technical requirement list, technical clause paragraph of project bidding document, and technical indicator paragraph of internal project establishment requirement specification; the technical supply document text set consists of supply-side source document entries, which include at least one of the following: patent specification text, paper abstract and method paragraph, technical standard clause paragraph, product manual key parameter paragraph, and scientific research achievement registration material; each document entry is processed as the smallest unit before entering the reading stage, and a unified document metadata field is written when the data is accessed. The document metadata field includes at least the following: document source type field, source site field, publication time field, and original link field.

[0023] Specifically, the document access module reads document entries one by one from the file system directory, database table, or result set returned by the interface, and generates a document identifier for each document entry; the text parsing module segments the main text of the document entries into paragraphs, and generates a fragment position identifier for each paragraph segmentation result, thereby obtaining a set of text fragments; the document identifier is a fixed-length string obtained by concatenating the source type + source site + original number + batch number into the database and then hashing it, which is used to stably refer to the same document entry in subsequent steps; the fragment position identifier is a combination of document identifier + paragraph number + sentence number + character start and end position identifier, which is used to trace the context fragments that appear in subsequent steps.

[0024] For example, a technical requirement document entry is “Technical requirements of a certain enterprise in 2025: to improve the accuracy of fault location and reduce manual inspection in low-voltage power distribution scenarios”, and its document identifier is DMD-REQ-202501-9F3A2C; when this document is segmented into the second sentence of the third paragraph, the segment position identifier is DMD-REQ-202501-9F3A2C#P03#S02#C015-C078.

[0025] Furthermore, the encoding normalization and symbol standardization are completed synchronously by the text parsing module when generating the text fragment set. The processing order is as follows: first, all text fragments are uniformly converted to the same character encoding; then, full-width / half-width symbols, different forms of hyphens, and different forms of quotation marks are uniformly replaced with the set standard symbols; then, consecutive whitespace characters are compressed into a single space; then, common unit expressions are standardized and mapped while retaining the reference fields of the original expressions.

[0026] S1.2. Perform dictionary matching extraction and sequence labeling model extraction on the text fragment set to obtain the first term candidate set and the second term candidate set. Then, filter and merge the first term candidate set and the second term candidate set according to the candidate boundary consistency rule to obtain the term candidate set.

[0027] In a preferred embodiment, dictionary matching extraction scans the normalized text field according to the rules of longest priority and conflict resolution at the same position to obtain a first term candidate set; each candidate entry in the first term candidate set is at least written with: candidate original text field, candidate start and end character position field, candidate source field and fragment position identifier, which are used to indicate which sub-table of the dictionary database the candidate is matched with.

[0028] In a preferred embodiment, the sequence labeling model extraction is performed by a trained sequence labeling model. The sequence labeling model takes the word segmentation sequence of the normalized text field as input and outputs a label sequence for each word. The label sequence distinguishes at least three types of labels: term start, term interior, and non-term. The sequence labeling model extracts a second term candidate set, and each candidate entry in the second term candidate set is written with: candidate original text field, candidate start and end character position field, candidate confidence field, and fragment position identifier, for subsequent filtering and merging.

[0029] Specifically, the candidate boundary consistency rule in this embodiment is used to filter and merge the first term candidate set and the second term candidate set. The merging order is as follows: first, the two sets are grouped by the fragment position identifier as the grouping key; then, within the same fragment position identifier group, the candidates are aligned according to the start and end character position fields; when the start and end character positions of the first candidate and the second candidate are completely consistent, one merged candidate is retained and written into the multi-source marker field as a dictionary + model; when the two only partially overlap, the candidate with the longer coverage is retained first, and an overlap conflict marker field is written into the retained candidate for subsequent verification; when only one candidate exists, the candidate is retained and written into the multi-source marker field as a dictionary; through the above filtering and merging, a term candidate set is obtained, so that term extraction retains the stability of the domain dictionary and absorbs the ability of the sequence labeling model to identify new words and variants.

[0030] S1.3 Generate normalized term identifiers for each term candidate in the term candidate set, and generate term type identifiers based on the character set and naming format of the term original text. Write the normalized term identifier, term type identifier, and corresponding term original text, document identifier, and fragment position identifier as fields to obtain the term entity set.

[0031] In this embodiment, the normalized term identifier is generated by the term normalization module. The generation process is as follows: first, case unification, whitespace compression, full-width and half-width character unification, and synonym abbreviation mapping are performed on the candidate original text fields; then, the normalized string is used as the normalized term key; when the normalized term key already exists in the term entity table, the existing normalized term identifier is reused; when the normalized term key does not exist, a new normalized term identifier is generated and written to the term entity table.

[0032] In a preferred embodiment, the terminology type identifier is generated based on the character set and naming format of the original terminology. The generation rules are as follows: when the original terminology contains obvious material / component suffixes and appears in the form of a noun, the type is written as component entity; when the original terminology contains methodological suffixes such as method, model, network, algorithm, etc., the type is written as method entity; when the original terminology contains unit, numerical range, or parameter name format, the type is written as parameter entity; when the original terminology is in the format of organization, standard number, or patent number, the type is written as identifier entity; and for those that cannot be clearly classified, the type is written as general terminology entity.

[0033] Each term entity record includes at least the following fields: normalized term identifier field, term type identifier field, term text field, document identifier field, fragment location identifier field, and frequency of occurrence field; wherein the frequency of occurrence field is the cumulative value of the number of times the same normalized term identifier appears in the same document identifier.

[0034] S1.4. Based on the original terminology and normalized terminology identifiers, calculate the string similarity value between demand terminology entities and supply terminology entities, and generate a set of candidate string pairs according to the string threshold and Top-M rule.

[0035] To reduce noise, candidate type threshold filtering is first performed based on term type identifiers. Only general term entities and method entities participate in string similarity calculation; parameter entities and identifier entities only participate in the calculation when they are completely identical. The string similarity calculation uses edit distance as a metric and maps it to the interval between zero and one. Before the calculation, synonym abbreviation mapping is performed on the term keys to make abbreviation relationships such as GNN and graph neural networks comparable at the string level. The calculation results are written to the string similarity value field and together with the demand normalized term identifier field and the supply normalized term identifier field, they form a candidate pair record.

[0036] In a preferred embodiment, the string threshold is the retention threshold of the string similarity value field, preferably 0.75; the Top-M rule is as follows: for each demand normalized term identifier, sort all supply normalized term identifiers in descending order by the string similarity value field and extract the top M to form a set of string candidate pairs, preferably M is 20; when the number of candidates for a demand term is less than M, retain all candidates that meet the string threshold.

[0037] It should be noted that this embodiment, by simultaneously employing threshold and Top-M rules, can control the computational scale when there are many demand terms and avoid missed recall when there are few demand terms.

[0038] S1.5 Based on the context fragments corresponding to the fragment position identifiers, generate semantic vectors for each term entity and calculate semantic similarity values. Generate a set of semantic candidate pairs according to the semantic threshold and the Top-N rule.

[0039] The text fragments are obtained by extending a preset length of characters forward and backward from the location where the term appears, and also include the paragraph title or clause title to which the fragment belongs as a supplementary context field; the semantic vector is obtained by encoding the context fragments using a text encoding model, and is written into the semantic vector field and stored in association with the normalized term identifier field, so that the semantic vector of the same term entity can be reused in subsequent steps without repeated encoding.

[0040] In a preferred embodiment, the semantic similarity value is obtained by measuring the vector similarity between the semantic vector of the demand term entity and the semantic vector of the supply term entity, and the calculation result is written into the semantic similarity value field; the semantic threshold is preferably 0.70; the Top-N rule is: for each demand normalized term identifier, sort all supply normalized term identifiers in descending order according to the semantic similarity value field and extract the top N to form a set of semantic candidate pairs, preferably N is 30; by introducing the title field and clause field in the context fragment, mismatches caused by terms with the same name but different meanings can be reduced.

[0041] S1.6 Merge the string candidate pair set and the semantic candidate pair set, and generate gating rules based on the term type identifier. Write the candidate pair set when the string similarity value is greater than or equal to the string threshold and the semantic similarity value is greater than or equal to the semantic threshold.

[0042] In a preferred embodiment, the string candidate pair set and the semantic candidate pair set are merged to remove duplicates using the demand normalized term identifier field + the supply normalized term identifier field as the joint key, and the merged set forms an initial candidate pair set. For each merged candidate pair record, both the string similarity value field and the semantic similarity value field are written. When a candidate pair comes from only one set, the other similarity field is written with a default value and the source field is marked for subsequent gating rule determination.

[0043] Furthermore, the gating rules are generated driven by term type identifiers, and the gating rules include at least the following three categories: First, when both the demand term type identifier and the supply term type identifier are method entities, the semantic similarity value field is required to reach a higher threshold, while the string similarity value field is allowed to be slightly lower; Second, when both are general term entities, the string similarity value field and the semantic similarity value field are required to reach the threshold simultaneously; Third, when either one is a parameter entity or an identifier entity, the string similarity value field is required to reach a higher threshold and meet the numbering consistency requirement.

[0044] It should be noted that the technical requirement text and the technical supply text are heterogeneous in origin and have significantly different expressions. Using only a single dictionary for extraction may easily miss emerging terms and variant expressions, and using only a single model for extraction may easily introduce domain noise. Therefore, dictionary matching extraction provides stable basic recall, sequence labeling model extraction is used to supplement new words and long terms, and candidate boundary consistency rules are used to resolve conflicts and maintain traceable boundaries, thereby obtaining a complete and traceable set of term entities.

[0045] Furthermore, candidate pair generation simultaneously incorporates string similarity and semantic similarity and applies gating rules, which can achieve a balance between two common errors: similar in form but different in meaning and similar in meaning but different in form. It also uses differentiated thresholds to reduce false graph matching when there are differences in term types. Compared with existing schemes that rely solely on keyword overlap or vector nearest neighbors, this embodiment writes document identifiers, fragment position identifiers, candidate pair identifiers, and rule version numbers at the candidate stage, reducing the cost of manual review in subsequent graph construction.

[0046] S2. Construct a candidate synonym cluster graph based on the candidate pair set and write it into the heterogeneous demand matching graph. Input the heterogeneous demand matching graph into a Bayesian graph neural network to obtain node embedding vectors, node uncertainty metric parameters, and edge uncertainty metric parameters. Note that the following should be noted in this step:

[0047] S2.1. Based on the candidate pair set, retrieve the demand term entities and supply term entities from the term entity set, and write the candidate pair identifier and similarity field for each candidate pair to obtain the candidate pair set with fields.

[0048] In this embodiment, the retrieval is completed using an index retrieval method. When the term entity set is written, an index has been created for the normalized term identifier field. Therefore, for each record in the candidate pair set, its required normalized term identifier field and supply normalized term identifier field are read first. Then, the corresponding term entity record is retrieved from the term entity set, and the retrieved term type identifier field, term original text field, document identifier field, and fragment position identifier field are written into the candidate pair extension field, thereby forming a candidate pair set with fields.

[0049] In a preferred embodiment, the candidate pair identifier is obtained by concatenating the demand normalized term identifier field, the supply normalized term identifier field, and the gating rule version number field and then performing hash encoding, which is used to reference the same candidate pair in the graph; the similarity field includes at least a string similarity value field and a semantic similarity value field, and retains the gating pass tag field and the source field.

[0050] S2.2. Generate connected components from the candidate pair set with fields, and mark the term entities in the same connected component as the same cluster identifier to obtain the candidate synonym cluster graph.

[0051] In this embodiment, a graph traversal method is used to generate connected components. Specifically, term entities are treated as nodes, and candidate pairs with fields are added to a temporary candidate graph as undirected connections. Then, the temporary candidate graph is traversed, and the set of reachable nodes is collected starting from any unvisited node. The set of reachable nodes constitutes a connected component. A cluster identifier is generated for each connected component and written into all term entity records within that component. The cluster identifier is preferably a CLUST- prefix plus a component sequence number or hash value to ensure uniqueness within the same graph version.

[0052] Preferably, by using connected components as the initial partition of synonymous clusters, many-to-many candidate pair relationships can be compressed into a cluster-level structure, providing a structural basis for the subsequent edge typification and consistency constraints of heterogeneous graphs.

[0053] S2.3 Write the candidate synonymous cluster graph into the heterogeneous demand matching graph, and write the intra-cluster synonymous edges and cross-cluster related edges according to the edge type rules of intra-cluster edges and cross-cluster edges. Intra-cluster synonymous edges correspond to candidate pairs with the same cluster identifier and store the candidate pair identifiers together. Cross-cluster related edges correspond to candidate pairs with different cluster identifiers and store the similarity field together.

[0054] In a preferred embodiment, the heterogeneous demand matching graph is a graph data structure containing multiple types of nodes and multiple types of edges. The nodes include at least: demand term nodes, supply term nodes, and cluster nodes; the edges include at least: intra-cluster synonym edges, cross-cluster association edges, and term-cluster affiliation edges; the graph is stored in a node table + edge table format when it is written. Each node record includes a node identifier field, a node type field, a normalized term identifier field or a cluster identifier field, a term type identifier field, and a source field; each edge record includes an edge identifier field, an edge type field, a start point identifier field, an end point identifier field, and an attribute field matching the edge type.

[0055] In a preferred embodiment, the edge typification rules include at least the following: when the term entities at both ends of a candidate pair have the same cluster identifier, the candidate pair is mapped to an intra-cluster synonymous edge, and the candidate pair identifier is associated and stored in the attribute field of the intra-cluster synonymous edge; when the term entities at both ends of a candidate pair have different cluster identifiers, the candidate pair is mapped to a cross-cluster association edge, and the string similarity value field and the semantic similarity value field are associated and stored in the attribute field of the cross-cluster association edge; at the same time, in order to make the cluster structure explicit, a term-cluster affiliation edge is written between each term entity and its cluster identifier, which is used to quickly locate the set of nodes within the cluster when applying consistency constraints within the cluster later.

[0056] Preferably, through the above writing method, the graph retains the similarity of candidate pairs and forms a structured input that can be used for propagation in graph neural networks.

[0057] S2.4. Generate node feature matrices and edge feature matrices based on heterogeneous demand matching graphs, and write the node feature matrices and edge feature matrices into the graph calculation input set.

[0058] In a preferred embodiment, the node feature matrix is ​​obtained by stacking the feature vectors of each node in the order of the node identifier fields; for term nodes, the node feature vector includes at least: a term semantic vector field, a category encoding field for term type identifier, a source type encoding field, and an occurrence frequency field; for cluster nodes, the node feature vector includes at least: an aggregation result field of semantic vectors of term nodes within the cluster, a cluster size field, and a cluster edge density field.

[0059] In a preferred embodiment, the edge feature matrix is ​​obtained by stacking the feature vectors of each edge in the order of the edge identifier fields; for cross-cluster association edges, the edge feature vector includes at least: a string similarity value field, a semantic similarity value field, a gating rule version number field, and a source tag field; for intra-cluster synonym edges, the edge feature vector includes at least: a candidate pair identifier reference field and an intra-cluster consistency initial weight field; for term-cluster affiliation edges, the edge feature vector includes at least: an affiliation strength field and a category encoding field of the cluster identifier field.

[0060] S2.5. Perform forward propagation operation of Bayesian graph neural network on the graph computation input set to obtain node embedding vector, node uncertainty metric parameter and edge uncertainty metric parameter.

[0061] Specifically, the Bayesian graph neural network aggregates the neighboring node information of the target node in each layer and outputs the node embedding vector, while providing measurement parameters for the uncertainty of node embedding and the uncertainty of edge propagation.

[0062] As an example, the mathematical expressions for node information aggregation and node embedding update are as follows:

[0063]

[0064] As an example, the mathematical expressions for the edge uncertainty metric and the edge weight are as follows:

[0065]

[0066] As an example, the mathematical expression for the nodal uncertainty metric parameter is:

[0067]

[0068] in, For nodes In the Update the node embedding vector of the layer; For nodes The set of adjacent neighboring nodes; For adjacent node indexes; For the first Layer from node To the node The adjacent edge weight coefficient; For the first The trainable linear transformation matrix of the layer; For the first Layer bias vector; It is a non-linear activation function; For the normalization operator on the adjacency set dimension; For the first Uncertainty sensitivity coefficient of the layer; For the first Layer edge The edge uncertainty metric parameter; For the first Layer nodes The node uncertainty metric parameter; For the first Layer nodes Update the node uncertainty metric parameters; For the first The layer is a coefficient used to adjust the cumulative magnitude of uncertainty; An operator for measuring the degree of discreteness of the embedding set of adjacent nodes; This is a truncation operator that limits values ​​to a preset range to avoid the unbounded growth of uncertain parameters.

[0069] It should be noted that candidate pairs come from both string and semantic perspectives, exhibiting structural features of cross-document, multi-source, and chain-like associations. If only table-based nearest neighbor methods are used, it is difficult to propagate evidence in the global structure and to express the differences between intra-cluster synonyms and cross-cluster associations. By explicitly writing the cluster structure and introducing edge typification rules, the graph neural network can distinguish the contribution sources of intra-cluster edges and cross-cluster edges to the embedding update during propagation, and quantify the propagation reliability with an uncertainty metric parameter, so that subsequent matching ranking not only reflects the degree of similarity, but also the consistency of evidence and the stability of propagation. Compared with existing schemes that only output a single similarity score, this embodiment retains the candidate pair identifier, similarity field, and rule version number in the graph structure, providing a cluster-level structural basis for the consistency constraint update of S3.

[0070] S3. Weighted aggregation of adjacent edge propagation information vectors based on edge uncertainty metric parameters, and application of consistency constraints within the candidate synonymous cluster graph to update node embedding vectors and node uncertainty metric parameters. Note the following in this step:

[0071] S3.1 Generate edge weight coefficients based on edge uncertainty measurement parameters, and associate the edge weight coefficients with the corresponding adjacent edge identifiers and write them into the edge weight set.

[0072] In a preferred embodiment, the edge weight coefficient is obtained by mapping the edge uncertainty metric parameter. The mapping rule is as follows: the larger the edge uncertainty metric parameter, the smaller the edge weight coefficient; the smaller the edge uncertainty metric parameter, the larger the edge weight coefficient. A preset segmented mapping rule is used during mapping to avoid extreme values ​​dominating the aggregation result: when the edge uncertainty metric parameter falls into the low uncertainty range, the edge weight coefficient is set to the high weight level; when the edge uncertainty metric parameter falls into the medium uncertainty range, the edge weight coefficient is set to the medium weight level; when the edge uncertainty metric parameter falls into the high uncertainty range, the edge weight coefficient is set to the low weight level.

[0073] Specifically, the boundaries of each interval are determined by the quantiles of the edge uncertainty metric parameter obtained from offline statistics. The distribution statistics of this edge uncertainty metric parameter use historical sample sets as the statistical sample source. The uncertainty metric parameter values ​​of all edges in the sample set are sorted and the quantiles are calculated. The following quantiles are preferably used as the two segment boundaries: the 30th percentile is used as the boundary between the low-level uncertainty interval and the medium-level uncertainty interval; the 70th percentile is used as the boundary between the medium-level uncertainty interval and the high-level uncertainty interval.

[0074] Accordingly, the intervals are divided as follows: when the edge uncertainty metric is less than or equal to the 30th percentile, it is determined to be a low-level uncertainty interval, and the edge weight coefficient is set to the high weight level of 0.80; when the edge uncertainty metric is greater than the 30th percentile and less than or equal to the 70th percentile, it is determined to be a medium-level uncertainty interval, and the edge weight coefficient is set to the medium weight level of 0.50; when the edge uncertainty metric is greater than the 70th percentile, it is determined to be a high-level uncertainty interval, and the edge weight coefficient is set to the low weight level of 0.20.

[0075] Furthermore, the edge weight set is written with the adjacent edge identifier as the index key. Each record includes at least: an adjacent edge identifier field, an edge weight coefficient field, an edge type field, and an edge uncertainty measurement parameter field, and a generation batch number field is written to associate it with the graph version identifier. Through this writing method, when performing adjacent edge aggregation on any target node in the future, the edge weight coefficient can be read directly by the adjacent edge identifier.

[0076] S3.2 For each target node in the heterogeneous demand matching graph, extract the adjacent edge information vectors that are adjacent to the target node, and perform weighted aggregation on the adjacent edge information vectors according to the edge weight set to obtain the aggregated representation vector.

[0077] In a preferred embodiment, the target node is a term node participating in the matching calculation in a heterogeneous demand matching graph. The term node includes at least demand term nodes and supply term nodes. For each target node, the set of adjacent edge identifiers adjacent to the target node is first read through the adjacency index of the graph. Then, the adjacent node embedding vector corresponding to each adjacent edge is read in sequence. The message representation after the adjacent node embedding vector is mapped by the edge type is used as the adjacent edge transmission information vector. In order to distinguish the semantics of different edge types for the message, the adjacent edge transmission information vector of intra-cluster synonymous edges preferably emphasizes synonym consistency information, the adjacent edge transmission information vector of cross-cluster related edges preferably emphasizes association similarity information, and the adjacent edge transmission information vector of term-cluster belonging edges preferably emphasizes cluster center representation information.

[0078] Specifically, weighted aggregation is performed by the set of edge weights: for each adjacent edge of the target node, the information vector is transmitted, the edge weight coefficient field corresponding to the adjacent edge identifier is read, and the edge weight coefficient is weighted and combined with the information vector transmitted by the adjacent edge to obtain the aggregated representation vector; the aggregation result is written to the temporary aggregation field of the target node and marked with the aggregation batch number so that it can be jointly updated with the node embedding vector in S3.3.

[0079] S3.3 Perform a weighted summation operation on the aggregated representation vector and the node embedding vector of the target node to obtain the updated node embedding vector, and write the updated node embedding vector back to the heterogeneous demand matching graph.

[0080] S3.4. Based on the cluster identifier of the candidate synonymous cluster graph, calculate the intra-cluster consistency deviation for the updated node embedding vector of the intra-cluster node, and perform consistency constraint update on the updated node embedding vector and node uncertainty metric parameter of the intra-cluster node according to the intra-cluster consistency deviation, to obtain the constrained updated node embedding vector and node uncertainty metric parameter.

[0081] In a preferred embodiment, the intra-cluster consistency deviation is calculated using the cluster identifier of the candidate synonymous cluster graph as the grouping key. For each cluster identifier, the set of nodes in the cluster and its updated node embedding vector are read first, and then the intra-cluster center representation vector is calculated. The intra-cluster center representation vector is the aggregation result of the updated node embedding vectors of the nodes in the cluster. Subsequently, the degree of deviation between the updated node embedding vector and the intra-cluster center representation vector of each node in the cluster is calculated, and the degree of deviation is written as the intra-cluster consistency deviation field. To avoid incomparability of deviations due to differences in cluster size, normalization is used when calculating the degree of deviation.

[0082] In a preferred embodiment, the consistency constraint update is performed according to the rule that the larger the deviation, the stronger the constraint. Specifically, when the intra-cluster consistency deviation of a node exceeds 0.15, the updated node embedding vector of the node is converged and corrected towards the direction of the intra-cluster center representation vector, and the node uncertainty metric parameter of the node is increased simultaneously. When the intra-cluster consistency deviation does not exceed 0.15, only a slight convergence correction is performed and the change of the node uncertainty metric parameter is kept within a small range. The update result is written to the node embedding vector field and the node uncertainty metric parameter field after the constraint update, respectively.

[0083] S4. Based on the updated node embedding vectors, generate matching ranking scores for demand and supply nodes. Then, according to the scoring threshold and Top-K rules, write the demand-supply matching edges into the heterogeneous demand matching graph. Simultaneously, write matching confidence and uncertainty measurement explanation fields into the demand-supply matching edges to generate an updated graph version and store it in the graph database. Note that the following points should be noted in this step:

[0084] S4.1 Calculate the embedding similarity value based on the node embedding vectors of demand nodes and supply nodes in the heterogeneous demand matching graph, and generate an uncertainty penalty value based on the node uncertainty measurement parameter and the edge uncertainty measurement parameter. Perform a synthesis operation on the embedding similarity value and the uncertainty penalty value to obtain the matching ranking score, where the synthesis operation is the embedding similarity value minus the uncertainty penalty value.

[0085] As an example, the mathematical expression for embedding similarity values ​​is:

[0086]

[0087] As an example, the mathematical expression for the uncertainty penalty value is:

[0088]

[0089] As an example, the mathematical expression for the matching ranking score is:

[0090]

[0091] in, Identify the demand node; Provide node identifiers; For demand nodes The node embedding vector after constraint update; For supply nodes The node embedding vector after constraint update; It is the inner product of two embedding vectors; The norm of the embedding vector for the demand node; The norm of the embedding vector for the supply node; To embed similarity values; For demand nodes The node uncertainty metric parameter; For supply nodes The node uncertainty metric parameter; For demand nodes With supply nodes The edge uncertainty metric parameter of the corresponding edges or the summation value of edge uncertainty obtained by mapping candidate edges; This is the weighting coefficient for demand-side uncertainty; This is the supply-side uncertainty weighting coefficient; The edge uncertainty weighting coefficient; ; This is a penalty value for uncertainty. Score the matching and ranking.

[0092] S4.2 Sort the candidate supply nodes of each demand node according to the matching ranking score, and filter them according to the scoring threshold and the Top-K rule to obtain the set of matching pairs to be entered into the graph. The scoring threshold is determined by the threshold parameter group, which includes at least the lower limit threshold of the matching ranking score and the upper limit threshold of the uncertainty penalty value. The candidate supply node is written into the set of matching pairs to be entered into the graph when the matching ranking score is greater than or equal to the lower limit threshold of the matching ranking score and the uncertainty penalty value is less than or equal to the upper limit threshold of the uncertainty penalty value. The Top-K rule is to select the first K elements (e.g., K is 10) in the sorting list corresponding to each demand node. When there are fewer than K candidates that meet the threshold conditions, all candidates that meet the threshold conditions are retained and the candidate shortage mark field is recorded.

[0093] In a preferred embodiment, when sorting the candidate supply nodes of each demand node according to the matching ranking score, the candidate supply node set is derived from the union of the candidate pair set formed in S1 and the cross-cluster association edge set written in S2, and a candidate list is established with the demand node identifier as the key. When sorting, the candidate supply nodes are sorted in descending order according to the matching ranking score. When the matching ranking scores are the same, they are sorted in ascending order according to the uncertainty penalty value, so that candidates with more stable evidence are given priority to enter the matching pair set in the graph.

[0094] S4.3. Generate demand-supply matching edges based on the set of matching pairs in the graph, and write the demand-supply matching edges into the heterogeneous demand matching graph. At the same time, write the matching confidence field and the uncertainty measurement interpretation field into the demand-supply matching edges.

[0095] In a preferred embodiment, when generating demand-supply matching edges based on the set of inbound matching pairs, each inbound matching pair is written as a matching edge into the edge table. During writing, a matching edge identifier is generated. The matching edge identifier is obtained by hashing the concatenation of the demand node identifier, the supply node identifier, and the graph version identifier to ensure uniqueness within the same graph version. After the matching edges are written, the edge types of the heterogeneous demand matching graph include at least intra-cluster synonym edges, cross-cluster association edges, term-cluster affiliation edges, and demand-supply matching edges, so that the candidate relationship and the final inbound relationship are structurally distinguishable.

[0096] The matching confidence field should include at least the following: matching ranking score field, embedding similarity value field, uncertainty penalty value field, ranking rank field, and Top-K hit tag field; the uncertainty measurement interpretation field should include at least the following: demand node uncertainty measurement parameter field, supply node uncertainty measurement parameter field, edge uncertainty measurement parameter field, gating rule version number field, and consistency constraint version number field.

[0097] S4.4 Generate a graph version identifier and associate the graph version identifier with the write record of the demand-supply matching edge to obtain the updated graph version.

[0098] In a preferred embodiment, the graph version identifier is generated by the version generation module and includes at least the following fields: data entry batch number, gating rule version number, threshold parameter group version number, model parameter version number, and generation timestamp. The version generation module concatenates and encodes the above fields to obtain the graph version identifier and writes it into the version table. At the same time, the graph version identifier is written back to the demand-supply matching edge record, node record, and edge record written in this instance, thereby forming a record-version association.

[0099] S4.5 Write the updated map version into the version storage area of ​​the map database according to the map version identifier.

[0100] In a preferred example of a technology demand scenario where talent and technological capability are matched, the set of technology demand documents includes enterprise job requirements and project team personnel requirements, while the set of technology supply documents includes candidate resumes, abstracts of talent research achievements, and patent texts.

[0101] In the example, the retrieved requirement document entry includes the job title: Intelligent Recommendation Algorithm Engineer; requirements: experience in graph neural network modeling, knowledge graph construction, and retrieval ranking optimization. Its document identifier is written as DMD-REQ-202601-0A19C4, and the fragment position identifier is written as DMD-REQ-202601-0A19C4#P02#S01#C001-C112. The retrieved supply document entry includes Candidate A: involved in graph neural network recommendation and knowledge graph projects; published papers on GNN and entity disambiguation; possesses experience in vector retrieval parameter tuning. Its document identifier is written as DMD-SUP-202601-71D2B8, and the fragment position identifier is written as DMD-SUP-202601-71D2B8#P01#S02#C010-C146.

[0102] After performing encoding normalization and symbol normalization on the text fragment set, the data is written into the normalized text field and the original text field, while retaining the index association between the document identifier and the fragment position identifier. Subsequently, dictionary matching extraction and sequence labeling model extraction are performed on the text fragment set to obtain a first term candidate set and a second term candidate set. These are then filtered and merged according to the candidate boundary consistency rule to obtain the final term candidate set. Specifically, term candidate graph neural networks, knowledge graphs, and retrieval ranking optimization are obtained from demand fragment extraction, while term candidate GNNs, knowledge graphs, vector retrieval parameter tuning, and entity elimination are obtained from supply fragment extraction. For each candidate term, a normalized term identifier and a term type identifier are generated and then written into the term entity set. For example, the normalized term identifier for graph neural networks is written as TERM-00004567, and the term type identifier is written as a method entity; the normalized term identifier for GNNs is written as TERM-00004568, and the term type identifier is written as a method entity; the normalized term identifier for knowledge graphs is written as TERM-00001210, and the term type identifier is written as a method entity. Each term entity record is simultaneously associated with the term original text field, document identifier, and fragment position identifier. The string similarity value is calculated based on the original terminology and normalized terminology identifiers, and a set of candidate string pairs is generated according to the string threshold and Top-M rule. In the example, the string threshold is set to 0.75 and the Top-M rule is set to retain the top 20 for each required term. The string similarity value field of the candidate pair Graph Neural Network-GNN is 0.80. Semantic vectors are generated based on context fragments, and semantic similarity values ​​are calculated. In the example, the semantic threshold is set to 0.70 and the Top-N rule is set to retain the top 30 for each required term. The semantic similarity value of the candidate pair Graph Neural Network-GNN is then obtained. The similarity value field is 0.90; after merging the string candidate pair set and the semantic candidate pair set, a gating rule is generated based on the term type identifier and written into the gating rule version number field as GR-0003. When the candidate pair meets the string retention threshold and the semantic retention threshold, it is written into the candidate pair set; in this embodiment, the string retention threshold is the same as the string threshold and is written as 0.75, and the semantic retention threshold is the same as the semantic threshold and is written as 0.70; the candidate pair that meets the conditions is written into the candidate pair identifier field, for example, the candidate pair identifier of the graph neural network-GNN is written as PAIR-7C21F0.

[0103] Furthermore, based on the candidate pair set, demand term entities and supply term entities are retrieved from the term entity set and written into the candidate pair set with fields. Then, connected components are generated based on the candidate pair set with fields and cluster identifiers are marked. In the example, the graph neural network and GNN fall into the same connected component, and the cluster identifier is written as CLUST-0008. When writing the candidate synonymous cluster graph into the heterogeneous demand matching graph, a node table and an edge table are written: the node table writes demand term nodes, supply term nodes, and cluster nodes, and the edge table writes intra-cluster synonymous edges and cross-cluster related edges according to the edge typification rules. The candidate pair identifier is associated and stored on the intra-cluster synonymous edges, and the string similarity value field and semantic similarity value field are associated and stored on the cross-cluster related edges. After generating the node feature matrix and edge feature matrix based on the heterogeneous demand matching graph and writing them into the graph calculation input set, the forward propagation operation of the Bayesian graph neural network is performed to obtain the node embedding vector. The parameters include fields for node uncertainty measurement and edge uncertainty measurement. In the example, the node uncertainty measurement parameter field for the demand term node graph neural network is written as 0.10, the node uncertainty measurement parameter field for the supply term node GNN is written as 0.08, and the edge uncertainty measurement parameter field for the cross-cluster association edge between the two is written as 0.06. Then, edge weight coefficients are generated based on the edge uncertainty measurement parameters and written into the edge weight set. The adjacent edge transmission information vectors of each target node are extracted and aggregated according to the edge weight set to obtain the aggregated representation vector. The aggregated representation vector is synthesized and updated with the node embedding vector of the target node and written back to the graph. Then, the intra-cluster consistency deviation is calculated according to the cluster identifier and the consistency constraint update is performed. In the example, the consistency deviation within the cluster identifier CLUST-0008 is in the low deviation range, and the node uncertainty measurement parameter field is maintained within 0.12.

[0104] During the graph update phase, embedding similarity values ​​are calculated based on the updated node embedding vectors. Uncertainty penalty values ​​are then generated by combining node uncertainty metrics and edge uncertainty metrics, resulting in a matching ranking score, which is written into the matching confidence field. In the example, the embedding similarity value for the demand node's graph neural network modeling experience and the supply node's GNN recommendation project experience is written as 0.78, the uncertainty penalty value as 0.09, and the matching ranking score as 0.69. Similarly, the embedding similarity value for the demand node's knowledge graph construction experience and the supply node's knowledge graph project experience is written as 0.74, the uncertainty penalty value as 0.11, and the matching ranking score as 0.63. Each demand node is sorted according to its matching ranking score, and the set of matching pairs to be included in the graph is selected based on a scoring threshold and Top-K rules. In the example, the lower threshold for the matching ranking score is written as 0.35, the upper threshold for the uncertainty penalty value is written as 0.25, and the Top-K rule is written as... The first 5 entries for each demand node are selected. For matching pairs that meet the criteria, demand-supply matching edges are generated and written into the heterogeneous demand matching graph. A matching confidence field and an uncertainty measurement explanation field are written to the demand-supply matching edges. The uncertainty measurement explanation field includes at least a summary of the values ​​of the demand node uncertainty measurement parameter field, the supply node uncertainty measurement parameter field, and the edge uncertainty measurement parameter field, as well as the gating rule version number field GR-0003. Finally, a graph version identifier is generated and associated with the current write record. In this example, the graph version identifier is written as VER-20260121-001. The updated graph version is then written to the version storage area of ​​the graph database according to the graph version identifier. In subsequent operations, the business side writes feedback on whether candidate A has entered the interview or been hired into a feedback table and associates it with the demand-supply matching edge identifier. The statistical results of the feedback table are written into the threshold parameter group version table, triggering the next round of graph version generation, thus forming an iterative closed loop based on version storage and feedback write-back.

[0105] In the application of the above embodiments, other aspects of the present invention also disclose a computer device, including: one or more processors and a memory.

[0106] The memory is used to store operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the graph neural network-based technical requirement matching graph construction method of the foregoing embodiments, especially... Figure 1 The flowchart of the method is shown.

[0107] Other aspects disclosed in the embodiments of the present invention also propose a computer-readable medium for storing software including instructions executable by one or more computers, which, upon execution, cause the one or more computers to perform operations including the flow of the graph neural network-based technical requirement matching graph construction method of the foregoing embodiments, particularly... Figure 1 The flowchart of the method is shown.

[0108] It should be recognized that embodiments of the present invention may be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium.

[0109] The method can be implemented using standard programming techniques, including a non-transitory computer-readable storage medium configured with a computer program in the computer program, wherein the storage medium is configured such that the computer operates in a specific and predefined manner.

[0110] Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system; however, if required, the program can be implemented in assembly or machine language.

[0111] In any case, the language can be either compiled or interpreted.

[0112] Furthermore, for this purpose, the program can run on programmed application-specific integrated circuits.

[0113] The processes described herein (or variations and / or combinations thereof) can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. The computer program includes a plurality of instructions executable by one or more processors.

[0114] Furthermore, the method can be implemented in any suitable computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices.

[0115] Various aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether portable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it can be read by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein.

[0116] Furthermore, machine-readable code, or parts thereof, can be transmitted via wired or wireless networks.

[0117] When such media includes instructions or programs that combine with a microprocessor or other data processor to implement the steps described above, the invention described herein includes these and other different types of non-transitory computer-readable storage media.

[0118] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for constructing a technical demand matching graph based on graph neural networks, characterized in that, include: Read the text sets of technical requirement documents and technical supply documents, perform term entity extraction to obtain a term entity set, and generate a candidate pair set based on string similarity and semantic similarity; Based on the candidate pair set, a candidate synonym cluster graph is constructed and written into the heterogeneous demand matching graph. The heterogeneous demand matching graph is then input into a Bayesian graph neural network to obtain node embedding vectors, node uncertainty measurement parameters, and edge uncertainty measurement parameters. The adjacent edge information vectors are weighted and aggregated according to the edge uncertainty metric parameter, and a consistency constraint is applied within the cluster of the candidate synonym cluster graph to update the node embedding vector and the node uncertainty metric parameter. Based on the updated node embedding vector, a matching ranking score for demand nodes and supply nodes is generated. The demand-supply matching edge is written into the heterogeneous demand matching graph according to the scoring threshold and Top-K rule of the matching ranking score. At the same time, a matching confidence field and an uncertainty measurement explanation field are written into the demand-supply matching edge to generate an updated graph version and store it in the graph database. The step of generating a matching ranking score between demand nodes and supply nodes based on the updated node embedding vector includes: Based on the node embedding vectors of the demand nodes and supply nodes in the heterogeneous demand matching graph, an embedding similarity value is calculated, and an uncertainty penalty value is generated based on the node uncertainty measurement parameter and the edge uncertainty measurement parameter. A synthesis operation is performed on the embedding similarity value and the uncertainty penalty value to obtain a matching ranking score, wherein the synthesis operation is the embedding similarity value minus the uncertainty penalty value. Candidate supply nodes for each demand node are sorted according to the matching ranking score, and a set of matching pairs for entering the graph is obtained by filtering according to the scoring threshold and the Top-K rule. The scoring threshold is determined by a threshold parameter group, which includes at least a lower limit threshold for matching ranking score and an upper limit threshold for uncertainty penalty value. Candidate supply nodes are written into the set of matching pairs for entering the graph when the matching ranking score is greater than or equal to the lower limit threshold for matching ranking score and the uncertainty penalty value is less than or equal to the upper limit threshold for uncertainty penalty value. The Top-K rule is to select the top K elements in the sorting list corresponding to each demand node.

2. The method for constructing a technical requirement matching graph based on graph neural networks according to claim 1, characterized in that, The method for extracting a set of term entities by performing term entity extraction includes: The read set of technical requirement documents and the set of technical supply documents are subjected to encoding normalization and symbol normalization processing, and document identifiers and fragment position identifiers are written to obtain a set of text fragments; Dictionary matching extraction and sequence labeling model extraction are performed on the text fragment set respectively to obtain a first term candidate set and a second term candidate set. The first term candidate set and the second term candidate set are then filtered and merged according to the candidate boundary consistency rule to obtain a term candidate set. A normalized term identifier is generated for each term candidate in the term candidate set, and a term type identifier is generated based on the character set and naming pattern of the term original text. The normalized term identifier, the term type identifier, the corresponding term original text, the document identifier, and the fragment position identifier are written as fields to obtain the term entity set.

3. The method for constructing a technical requirement matching graph based on graph neural networks according to claim 2, characterized in that, The method for generating the candidate pair set includes: Based on the original terminology and the normalized terminology identifier, calculate the string similarity value between the demand terminology entity and the supply terminology entity, and generate a set of candidate string pairs according to the string threshold and the Top-M rule; Based on the context fragments corresponding to the fragment position identifiers, semantic vectors of each term entity are generated and semantic similarity values ​​are calculated. A set of semantic candidate pairs is generated according to the semantic threshold and the Top-N rule. The set of candidate strings and the set of candidate semantic pairs are merged, and gating rules are generated according to the term type identifier. Candidate pairs are written into the candidate pair set when the string similarity value is greater than or equal to the string threshold and the semantic similarity value is greater than or equal to the semantic threshold.

4. The method for constructing a technical requirement matching graph based on a graph neural network according to claim 3, characterized in that, Constructing a candidate synonym cluster graph based on the candidate pair set and writing it into the heterogeneous demand matching graph includes: Based on the candidate pair set, demand term entities and supply term entities are retrieved from the term entity set, and candidate pair identifier and similarity field are written for each candidate pair to obtain a candidate pair set with fields. Based on the candidate pair set with fields, connected components are generated, and term entities within the same connected component are marked as the same cluster identifier to obtain a candidate synonym cluster graph; Write the candidate synonym cluster graph into the heterogeneous demand matching graph, and write intra-cluster synonym edges and cross-cluster related edges according to the edge type rules of intra-cluster edges and cross-cluster edges. The intra-cluster synonym edges correspond to candidate pairs with the same cluster identifier and store the candidate pair identifiers together. The cross-cluster related edges correspond to candidate pairs with different cluster identifiers and store the similarity field together.

5. The method for constructing a technical requirement matching graph based on a graph neural network according to claim 4, characterized in that, The process of obtaining the node embedding vector, node uncertainty metric parameters, and edge uncertainty metric parameters includes: Based on the heterogeneous demand matching graph, a node feature matrix and an edge feature matrix are generated, and the node feature matrix and the edge feature matrix are written into the graph calculation input set; The forward propagation operation of a Bayesian graph neural network is performed on the graph computation input set to obtain node embedding vectors, node uncertainty metric parameters, and edge uncertainty metric parameters.

6. The method for constructing a technical requirement matching graph based on graph neural networks according to claim 5, characterized in that, Update methods include: The edge weight coefficients are generated based on the edge uncertainty metric parameters, and the edge weight coefficients are associated with the corresponding adjacent edge identifiers and written into the edge weight set. For each target node in the heterogeneous demand matching graph, the adjacent edge information vectors adjacent to the target node are extracted, and the adjacent edge information vectors are weighted and aggregated according to the edge weight set to obtain the aggregated representation vector. The aggregated representation vector and the node embedding vector of the target node are weighted and summed to obtain the updated node embedding vector, and the updated node embedding vector is written back to the heterogeneous demand matching graph. Based on the cluster identifier of the candidate synonymous cluster graph, the intra-cluster consistency deviation is calculated for the updated node embedding vector of the intra-cluster node, and consistency constraint update is performed on the updated node embedding vector and the node uncertainty metric parameter of the intra-cluster node according to the intra-cluster consistency deviation, so as to obtain the constrained updated node embedding vector and node uncertainty metric parameter.

7. The method for constructing a technical requirement matching graph based on graph neural networks according to claim 1, characterized in that, The generated updated map version includes: Based on the set of matching pairs in the graph, a demand-supply matching edge is generated, and the demand-supply matching edge is written into the heterogeneous demand matching graph. At the same time, a matching confidence field and an uncertainty measurement interpretation field are written onto the demand-supply matching edge. Generate a graph version identifier and associate the graph version identifier with the write record of the demand-supply matching edge to obtain the updated graph version; The updated map version is written into the version storage area of ​​the map database according to the map version identifier.

8. A computer device, characterized in that, include: One or more processors; The memory stores operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the technical requirement matching graph construction method based on any one of claims 1 to 7.

9. A computer-readable medium for storing software, characterized in that: The software includes instructions executable by one or more computers, which cause the one or more computers to perform operations, including the flow of the technical requirement matching graph construction method based on any one of claims 1 to 7.

Images