Project information automatic matching system for small and medium-sized enterprises

Abstract

The invention discloses an automatic matching system for project information of medium and small enterprises, which relates to the technical field of information processing, and comprises the following steps: generating enterprise feature sub-graphs in a field general knowledge graph according to field semantic tags of initial project information data; determining a node diffusion path according to the cross-layer propagation probability of nodes in the enterprise feature sub-graph, and generating a domain feature migration vector in the domain general knowledge graph according to the node diffusion path; constructing a multi-view reasoning network based on the domain feature migration vector, and determining an enterprise-project candidate matching group according to a node incidence matrix output by the multi-view reasoning network; constructing a cross-domain semantic reasoning network according to the semantic reconstruction probability of the enterprise-project candidate matching group in the domain general knowledge graph, and generating a target project information matching set according to the enterprise-project matching probability output by the cross-domain semantic reasoning network; according to the invention, accurate matching between the enterprise declaration intention and the project support condition based on technology internal association is realized.

Description

Technical Field

[0001] This invention relates to the field of information processing technology, specifically an automatic matching system for project information of small and medium-sized enterprises. Background Technology

[0002] With the continuous growth in the number of science and technology project applications, how to accurately and efficiently match enterprises' application intentions with project support conditions has become one of the key issues in the implementation of current industrial policies and resource allocation. Especially for small and medium-sized enterprises (SMEs) that lack project application experience and historical data support, due to the unstructured nature and cross-domain semantic ambiguity of their technical information, traditional manual review or keyword search methods suffer from insufficient matching accuracy and waste of resources, making it difficult to achieve truly accurate matching.

[0003] Meanwhile, currently widely used information matching technologies are mostly based on shallow textual semantics or single-modal data features, failing to fully explore the semantic relationships and feature transfer patterns of enterprise technical capabilities across different domains. Especially when enterprise business and project conditions involve multiple overlapping domains, existing methods lack effective cross-domain semantic fusion analysis mechanisms, resulting in poor actual adaptation between enterprises and projects, and seriously affecting the efficient allocation of project support resources. Summary of the Invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes an automatic matching system for project information of small and medium-sized enterprises.

[0005] To achieve the above objectives, the present invention provides an automatic matching system for project information of small and medium-sized enterprises, comprising: The subgraph generation module is used to obtain the initial project information data and domain general knowledge graph of the target enterprise, and generate enterprise feature subgraphs in the domain general knowledge graph based on the domain semantic tags of the initial project information data. The feature acquisition module is used to determine the node diffusion path based on the cross-layer propagation probability of nodes in the enterprise feature subgraph, and generate domain feature transfer vectors in the domain general knowledge graph based on the node diffusion path. The data analysis module is used to construct a multi-view graph inference network based on domain feature transfer vectors, and to determine enterprise-project candidate matching groups based on the node association matrix output by the multi-view graph inference network. The information matching module is used to construct a cross-domain semantic reasoning network based on the semantic reconstruction probability of enterprise-project candidate matching groups in the domain general knowledge graph, and generate a target project information matching set based on the enterprise-project matching probability output by the cross-domain semantic reasoning network.

[0006] Compared with the prior art, the beneficial effects of the present invention are: This invention introduces a multi-dimensional semantic analysis and structured modeling mechanism for enterprise technical information, which can accurately depict the enterprise's true technical capabilities and potential development direction without relying on historical application data. This effectively solves the problem that the technical characteristics of SMEs are difficult to accurately identify in the first application or cross-domain application scenarios, and improves the completeness and objectivity of the enterprise's application intention.

[0007] This invention, by constructing a form of expression for the dissemination and transfer of enterprise technical characteristics in the knowledge structure, enables in-depth exploration of the correlation between enterprise technical capabilities at different technical levels and related fields. This allows the matching between project support conditions and enterprise technical capabilities to go beyond superficial semantic consistency and instead make comprehensive judgments based on the intrinsic technical correlations, thereby improving the accuracy and rationality of matching between enterprises and projects.

[0008] This invention uses multi-perspective semantic reasoning and cross-domain semantic adaptation mechanisms to systematically evaluate the complex relationship between enterprises' application intentions and project support conditions. Under the premise of ensuring that the calculation process is controllable and feasible, it realizes intelligent adaptation between enterprises and projects, effectively improving the efficiency of resource allocation and overall implementation effect of science and technology projects. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.

[0010] Figure 1 This is a block diagram of the system of the present invention. Detailed Implementation

[0011] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] Please see Figure 1 This embodiment provides an automatic matching system for project information of small and medium-sized enterprises, including: The subgraph generation module 101 is used to obtain the initial project information data and domain general knowledge graph of the target enterprise, and generate an enterprise feature subgraph in the domain general knowledge graph based on the domain semantic tags of the initial project information data. In one specific embodiment, the target enterprise can be a small or medium-sized enterprise (SME) applying for a science and technology project for the first time and lacking historical project information, such as a technology company mainly engaged in the research and development of intelligent agricultural drones. Initial project information data is preferably obtained through automatic crawling of publicly available channels such as the company's official website and business registration information platforms, mainly including descriptions of the company's business scope, core technology specifications, textual descriptions of products and services, and image data.

[0013] It should be noted that the general knowledge graph of the domain is a pre-constructed structured semantic knowledge network. The nodes in the graph represent specific technical fields or technical concepts (e.g., "smart agriculture", "agricultural monitoring", "unmanned plant protection", "precision application", etc.), and the edges between nodes represent semantic relationships, which are usually constructed and obtained in advance by domain experts.

[0014] Specifically, the process of generating the enterprise feature sub-graph is as follows: Multimodal semantic features are extracted from the company's initial project information data, and the multimodal semantic features are divided into several semantic subspaces; Specifically, the multimodal semantic features in this embodiment include two aspects: text modality features and visual modality features. In one specific embodiment, the extraction of text modality features preferably uses a domestically developed pre-trained language model (such as the Baidu ERNIE model) to semantically encode the textual information provided by the enterprise (such as the enterprise's business scope, product technology introduction, etc.) into feature vectors; the extraction of visual modality features preferably uses a domestically developed visual feature model (such as the ResNet-50 model under the Baidu PaddlePaddle framework) to convert the enterprise's product image information into feature vectors.

[0015] Furthermore, in this embodiment, the multimodal feature vectors are preferably reduced using a variational autoencoder (VAE) to facilitate efficient computation and subsequent analysis. For example, the initial 768-dimensional text feature vector and 2048-dimensional visual feature vector are reduced to a unified 256-dimensional feature space, thereby obtaining a unified low-dimensional feature representation.

[0016] In one specific embodiment, this embodiment uses a clustering algorithm (such as the K-means algorithm) to perform clustering analysis on the dimensionality-reduced unified feature representation, and for example divides it into 3 independent semantic subspaces to clearly express the multidimensional technical characteristics of the enterprise. For example, it can be divided into "agricultural technology semantic subspace", "UAV flight control technology semantic subspace" and "automation technology semantic subspace".

[0017] For each semantic subspace, calculate the cross-semantic distance with nodes in the domain-general knowledge graph, and obtain domain semantic labels through clustering using the semantic distance function; Specifically, in this embodiment, the method for calculating the cross semantic distance preferably uses cosine distance to accurately measure the degree of semantic matching between feature vectors and knowledge graph nodes within the semantic subspace, thereby achieving accurate selection of domain semantic labels.

[0018] For example, in the "UAV flight control technology semantic subspace," the cross-semantic distance is calculated between the feature vector of the enterprise in this subspace and the feature vector of each node in the domain knowledge graph. Then, cluster analysis is performed based on the calculated semantic distance matrix, preferably using a spectral clustering algorithm to determine the node groups with high semantic matching as domain semantic labels. For example, if the knowledge graph nodes "UAV," "precision pesticide application," and "agricultural monitoring" form a class in the clustering results, these nodes can be identified as the domain semantic labels for this semantic subspace.

[0019] Using domain semantic tags as node activation conditions, and iteratively filtering based on the interaction probability distribution of node activation in the domain general knowledge graph, enterprise feature subgraphs are generated. Specifically, in this embodiment, the knowledge graph nodes corresponding to the domain semantic tags are used as the initial activated nodes. Then, based on the interaction activation probabilities between nodes, the system iteratively expands to determine the specific node set of the enterprise feature subgraph. The node activation probability is defined as the probability that the current node propagates semantic information to its neighboring nodes, preferably using a softmax normalization calculation method based on node semantic similarity. An exemplary calculation process is as follows: Assuming the currently active node is node The set of neighboring nodes is Then the node Activate neighbor nodes The probability is: In the formula, This refers to the semantic similarity between nodes (e.g., cosine similarity). This is a probability adjustment coefficient, determined based on experimental data, for example, a value of 0.5.

[0020] Furthermore, node activation expansion is performed iteratively, and the specific process includes: First, using the nodes corresponding to the defined domain semantic labels as the initial activation nodes, calculate their activation probabilities with their neighboring nodes; Then, neighboring nodes whose activation probability meets the preset threshold condition (e.g., activation probability not less than 0.05) are included in the enterprise feature subgraph, and the newly added nodes are used as the starting nodes of the next round of activation process, and the above process continues. For example, the neighboring nodes of the initially activated node "drone" "precision application", "agricultural monitoring" and "pest and disease identification" have activation probabilities of 0.12, 0.09 and 0.03, respectively. At this time, the first two nodes enter the feature subgraph, and the third node is not activated because its probability does not reach the threshold. Repeat the above iterative process until the termination condition is met, that is, the activation probability of all nodes to be activated is lower than the threshold (e.g., 0.05), or the maximum number of iterations is reached (e.g., 3 times), then the activation expansion is terminated.

[0021] The feature acquisition module 102 is used to determine the node diffusion path based on the cross-layer propagation probability of nodes in the enterprise feature subgraph, and generate a domain feature transfer vector in the domain general knowledge graph based on the node diffusion path. In one specific embodiment, the cross-layer propagation probability of nodes in the enterprise feature subgraph is used to represent the ability of a specific node to propagate information or influence semantics on other nodes in the graph network. The higher the propagation probability, the greater the contribution of the node to the subsequent matching process. Therefore, it is necessary to determine the effective node diffusion path based on the propagation probability.

[0022] Specifically, determining the node diffusion path based on the cross-layer propagation probability of nodes in the enterprise feature subgraph includes: Construct a dynamic game model for propagation between nodes in the enterprise feature subgraph, and determine the initial strategy distribution for node propagation based on the dynamic game model; Specifically, this embodiment views the enterprise feature subgraph as a propagation game network composed of multiple nodes, where each node acts as a player in the game and interacts with its neighbors. A payoff function for node propagation is defined to characterize the payoff for information propagation between nodes. The specific payoff function definition process is as follows: Assume nodes in the enterprise feature subgraph The set of neighboring nodes is ,node The dissemination strategy used Its value is the probability that a node participates in propagation (e.g., the range is...). The node reward function is: In the formula, For nodes and nodes The interaction benefit coefficient between nodes is obtained by calculating the semantic similarity between nodes; For nodes The propagation cost coefficient is determined based on experimental data, and can be 0.1 for example.

[0023] Furthermore, by solving for the Nash equilibrium of the above dynamic game model, the initial policy probability distribution of each node can be obtained. Specifically, the optimal response dynamic algorithm can be used to determine the initial policy distribution. .

[0024] The node diffusion stability state of the initial policy distribution is evaluated using the policy evolution stability condition to obtain the initial cross-layer propagation probability; Specifically, this embodiment uses the policy evolution stability condition to determine the stable state of node propagation, and defines the node policy evolution dynamic equation (replication dynamic equation) as follows: ;in, Represents a node The strategy probability, For nodes The benefits, This is the average of the profits of all nodes, i.e.: .

[0025] Subsequently, the above dynamic equations are solved using numerical iterative methods (such as the Runge-Kutta method). When the change in the propagation strategy probability of all nodes satisfies the stability condition (e.g., the change is less than 0.001), the cross-layer propagation probability distribution of nodes in the steady state is obtained. .

[0026] Under stable propagation conditions, the node diffusion strategy is iteratively updated through the path game benefit function until the node diffusion strategy converges, thus determining the node diffusion path. Specifically, the path game benefit function is used to comprehensively evaluate the benefits of joint information propagation among nodes in a path. The definition process of the path game benefit function is as follows: Assuming the set of path nodes is P, the path game payoff function is: In the formula, For nodes The synergistic benefit coefficient of inter-node joint propagation is preferably determined by the semantic similarity of nodes. For nodes The additional propagation cost in the path, determined based on experimental data, is exemplarily set to 0.1.

[0027] Subsequently, the gradient ascent algorithm is used to iteratively optimize the path game benefit function and update the node diffusion strategy: ;in, Let be the node diffusion strategy for the t-th iteration. The update rate is determined based on experimental data; an example value of 0.05 is used.

[0028] Through multiple iterations, when the update magnitude of the node diffusion strategy is less than the convergence threshold (e.g., 0.0001), the node diffusion strategy is considered to have reached convergence. Subsequently, the diffusion path is determined according to the node diffusion probability from large to small. For example, the key diffusion path is formed by selecting the node sequence with a probability greater than a preset threshold (e.g., "UAV → Agricultural Monitoring → Precision Pesticide Application").

[0029] Specifically, the step of generating domain feature transfer vectors in the domain-wide knowledge graph based on node diffusion paths includes: The semantic embedding features of each node in the node diffusion path are aggregated through a multi-layer graph attention mechanism to generate a path embedding representation; It should be understood that the node diffusion path reflects the propagation trajectory of an enterprise's core capabilities within the domain knowledge graph, while path embedding is used to perform overall semantic modeling of this propagation trajectory. In a specific embodiment, let the node diffusion path be... , where each node Each corresponds to a semantic embedding vector .

[0030] Specifically, this embodiment employs a multi-layer graph attention mechanism to weighted aggregate the features of nodes in the path. In each layer, nodes... Its neighboring nodes The attention weights are adaptively learned from the correlations between node features and then normalized to form an attention distribution. Through multiple iterations, nodes that contribute more to semantic propagation in the path are given higher weights in the aggregation result.

[0031] After multi-layer graph attention aggregation, the updated features of all nodes in the path are aggregated at the path level to obtain the path embedding representation. This is used to characterize the comprehensive semantic information carried by the diffusion path.

[0032] An embedding feature entropy weight allocation strategy is used to perform unbalanced optimization and adjustment on the semantic contribution of different nodes in the path embedding representation, resulting in an unbalanced optimized path embedding representation. It should be noted that the semantic contributions of nodes in the path vary significantly, and direct average aggregation can easily weaken the role of key nodes. Therefore, this embodiment introduces an embedded feature entropy weight allocation strategy to further optimize the path embedding representation.

[0033] Among them, an embedded feature entropy weight allocation strategy is used to perform unbalanced optimization and adjustment of the semantic contribution of different nodes in the path embedding representation, including: The initial entropy weight distribution of the embedded features of the nodes in the path embedding representation is computed, and the entropy weight distribution is adjusted with the difference in path propagation mode as a constraint. Specifically, the path embedding representation is decomposed into a set of node-level embedded features, and the information entropy value of each node feature in the path dimension is calculated. Nodes with lower information entropy represent more concentrated feature distribution, higher discriminative power, and more stable contribution to path semantics, and are therefore assigned higher weights in the initial entropy weight distribution.

[0034] Furthermore, the differences in path propagation patterns are introduced as constraints, and the propagation probability of a node in the diffusion path, its path position, and its cooperative relationship with other nodes are incorporated into the weight correction factor, so that the entropy weight distribution can simultaneously reflect the semantic information content and propagation influence of the node.

[0035] Based on the inter-node feature entropy weight interaction gain condition, the iterative update rate of the entropy weight distribution is dynamically adjusted until the entropy weight distribution meets the convergence condition. Specifically, this embodiment defines the entropy weight interaction gain between node pairs to measure the collaborative enhancement effect between nodes in path semantic modeling. If two nodes have high similarity in the semantic space and a high propagation probability, their interaction gain is large, correspondingly increasing the entropy weight update magnitude of both.

[0036] During the iteration process, the update rate of the entropy weight of each node is dynamically adjusted according to the entropy weight interaction gain, so that the weight distribution gradually stabilizes. When the change amplitude of the entropy weight of each node is less than the preset threshold in two consecutive iterations, the entropy weight distribution is considered to have reached a stable state.

[0037] By using the converged entropy weight distribution to redistribute the weights of the path embedding representation, an unbalanced optimized path embedding representation is obtained. Specifically, the converged node entropy weights are used as weight coefficients to perform a weighted summation of the node embedding features output by the multi-layer graph attention mechanism, resulting in the imbalanced optimized path embedding representation. The contribution of key nodes to path semantics is effectively amplified.

[0038] Based on the unbalanced optimized path embedding representation, a domain feature transfer vector with a multi-scale semantic association structure is generated. It is understandable that different semantic scales can reflect the migration characteristics of enterprise capabilities at different levels of abstraction. Based on the path embedding representation optimized by imbalance, this embodiment reorganizes features from multiple semantic scales, for example, by constructing multi-scale semantic representations through different receptive field lengths or feature grouping methods.

[0039] Subsequently, the semantic representations at various scales are fused to generate the final domain feature transfer vector. This transfer vector can simultaneously reflect the comprehensive capabilities of an enterprise at both the fine-grained technical level and the macro-domain level, providing a stable and information-dense input for the subsequent construction of a multi-view graph inference network.

[0040] Data analysis module 103 is used to construct a multi-view graph inference network based on domain feature transfer vectors, and to determine enterprise-project candidate matching groups based on the node association matrix output by the multi-view graph inference network. Specifically, the domain feature transfer vector reflects the cross-domain semantic propagation characteristics of the enterprise feature subgraph across different domains. In order to further achieve accurate matching between enterprises and projects, this embodiment constructs a multi-view graph inference network based on the transfer vector, so as to deeply mine and accurately characterize the relationship between enterprises and projects from multiple different analytical perspectives.

[0041] Specifically, the construction of the multi-view graph inference network based on domain feature transfer vectors includes: Multi-scale semantic segmentation is performed on the domain feature transfer vector to generate several sub-vectors with semantic granularity differences; Specifically, in this embodiment, the domain feature transfer vector is a multi-dimensional semantic representation generated in the aforementioned step S102. For example, its dimension is set to 768 dimensions to fully characterize the comprehensive characteristics of an enterprise in cross-domain semantic propagation. To achieve multi-scale analysis of the relationship between the enterprise and the project from different semantic granularities, preferably, the transfer vector is divided according to different granularities, for example, into three different granularity sub-vectors: small-scale (128 dimensions for example), medium-scale (256 dimensions for example), and large-scale (384 dimensions for example), denoted as... and This enables semantic analysis ranging from fine-grained to coarse-grained.

[0042] Based on the matching strength between the sub-vector granularity difference and the domain-general knowledge graph, a graph structure adaptive learning strategy is adopted to construct a multi-view initial sub-network; Specifically, in this embodiment, the semantic matching strength between each semantic granularity sub-vector and the nodes in the domain general knowledge graph is calculated respectively. Cosine similarity is preferably used as the metric for matching strength in order to construct a multi-view initial sub-network corresponding to each scale.

[0043] For example, using mesoscale subvectors For example, the semantic matching strength is calculated by comparing each subvector with the feature vectors of all nodes in the knowledge graph. After obtaining the matching strength distribution, a subvector is constructed. The initial subnetwork is centered around this core. In this network, the edge weights between nodes are determined by the matching strength, reflecting the degree of matching between different nodes for semantics at this scale.

[0044] Subsequently, similarly, small-scale sub-vectors were processed. and large-scale subvectors By performing the above construction process, multiple initial sub-networks with different granularities from multiple perspectives are obtained.

[0045] By leveraging the advantages of cross-granularity semantic propagation between sub-networks, the initial sub-networks are merged and the final multi-view graph reasoning network is constructed. Specifically, this embodiment defines a cross-granularity semantic propagation advantage coefficient to evaluate the semantic complementarity between sub-networks of different granularities, thereby achieving effective fusion of multiple initial sub-networks.

[0046] For example, the s-th subnetwork and the s-th subnetwork Dominance coefficient of cross-granular semantic propagation between subnetworks Defined as: In the formula, and Representing the s-th and the s-th respectively The edge weights from node i to node j in each subnetwork. A higher cross-granularity semantic propagation dominance coefficient indicates stronger semantic complementarity between the two networks.

[0047] Furthermore, based on the cross-granularity semantic propagation dominance coefficient, the fusion weight coefficient of each initial sub-network is determined. For example: In the formula, L represents the number of subnetworks (e.g., L=3).

[0048] Finally, by fusing weight coefficients, the sub-networks are weighted and fused to obtain the adjacency matrix of the final multi-view graph inference network. .

[0049] The above fusion results fully combine the semantic advantages of different scales, forming a more comprehensive and stable multi-view graph reasoning network structure.

[0050] Specifically, determining the enterprise-project candidate matching group based on the node association matrix output by the multi-view graph inference network includes: The propagation features of the node association matrix are mapped to a low-dimensional semantic space through probabilistic manifold learning to generate a semantically compact representation; Specifically, it is preferable to use a probabilistic manifold learning algorithm (such as the t-SNE algorithm) to map the high-dimensional node propagation features expressed by the node association matrix to a low-dimensional space, so as to obtain a more compact node representation that is easier to analyze later.

[0051] In its implementation, the mapping process of probabilistic manifold learning is achieved by minimizing the KL divergence of node similarity distributions in high-dimensional and low-dimensional spaces. The node representation in the low-dimensional space is denoted as... ,in Let be the low-dimensional feature representation of node i, and n be the number of nodes.

[0052] By leveraging the semantic similarity of nodes and frequent patterns of topological propagation in compact representations, several initial enterprise-project matching structures are formed. Specifically, this embodiment utilizes the aforementioned low-dimensional semantic compact representation to calculate the semantic similarity between enterprise nodes and project nodes to further determine the initial matching structure. Cosine similarity is preferably used for similarity, and the similarity of node pairs (u,v) is... The definition is as follows: In the formula, Let u and v represent the low-dimensional semantic representations of nodes u and v, respectively.

[0053] Furthermore, by analyzing the frequent patterns of topology propagation between nodes and combining the aforementioned semantic similarity index, the initial enterprise-project matching structure is determined. For example, when the semantic similarity of node pair (u,v) reaches or exceeds a set threshold (e.g., 0.7), node pair (u,v) is preliminarily determined as the initial matching structure.

[0054] Based on the semantic and topological coupling degree of the initial matching structure, candidate matching groups of enterprise-project with stable structural coupling are selected; Specifically, in this embodiment, the semantic and topological coupling degree Couple(u,v) is defined to evaluate the stability of the initial matching structure. An example of the calculation formula is as follows: In the formula, is the weighting coefficient between semantic similarity and topological structure, determined based on experimental data, and can be exemplarily taken as 0.5; N(u) and N(v) are the sets of neighboring nodes of nodes u and v in the graph reasoning network, respectively.

[0055] Furthermore, candidate matching groups are filtered based on the degree of structural coupling. For example, when When the value exceeds a set threshold (e.g., 0.6), the node pair (u,v) is determined to have a stable matching relationship and enters the final enterprise-project candidate matching group.

[0056] The information matching module 104 is used to construct a cross-domain semantic reasoning network based on the semantic reconstruction probability of enterprise-project candidate matching groups in the domain general knowledge graph, and generate a target project information matching set based on the enterprise-project matching probability output by the cross-domain semantic reasoning network. Specifically, the step of constructing a cross-domain semantic reasoning network based on the semantic reconstruction probability of enterprise-project candidate matching groups in the domain-general knowledge graph includes: Calculate the heterogeneous mapping relationship between the topological propagation structure and the semantic structure of the domain-general knowledge graph among candidate matching group node pairs, and identify key semantic reconstruction nodes; Specifically, this embodiment defines a heterogeneous mapping relationship between the topological propagation structure and the semantic structure of the domain knowledge graph to identify key nodes that significantly promote cross-domain matching. The specific heterogeneous mapping strength... The calculation formula is: In the formula, Indicates enterprise-project candidate node pair The set of local topological paths in the candidate matching group; This represents the set of semantic paths for the corresponding node pair in the domain-wide knowledge graph.

[0057] Subsequently, this embodiment sets a heterogeneous mapping intensity threshold (e.g., 0.5) and identifies nodes that meet the threshold condition as key semantic reconstruction nodes to support the optimization of subsequent cross-domain semantic propagation paths.

[0058] Based on the cross-domain semantic reconstruction strength of key semantic reconstruction nodes, the initial cross-domain semantic propagation path is dynamically optimized and adjusted; Specifically, defining the intensity of cross-domain semantic reconstruction. The formula for calculating the semantic reconstruction capability of key nodes is as follows: ,in, Represents a node The set of neighboring nodes; Representing neighboring nodes The weighting coefficient can be set to 0.5 for the initial example.

[0059] Furthermore, this embodiment dynamically optimizes and adjusts the initial cross-domain semantic propagation path, specifically as follows: If the cross-domain semantic reconstruction strength of key nodes If the value exceeds a set threshold (e.g., 0.6), the node is dynamically added to the semantic propagation path to form an optimized cross-domain semantic propagation path. .

[0060] Based on the propagation entropy of the optimized semantic propagation path, a cross-domain semantic reasoning network is constructed; Specifically, this embodiment defines propagation entropy. This is used to measure the stability of semantic propagation within a path. Propagation entropy. The calculation method is as follows: Among them, the propagation probability of nodes in the path The calculation is as follows: ; Furthermore, by setting a propagation entropy threshold (e.g., 0.3), paths with stable propagation entropy are selected, and the final cross-domain semantic reasoning network is constructed for subsequent calculation of accurate matching probability.

[0061] Specifically, the step of generating a target project information matching set based on the enterprise-project matching probability output by the cross-domain semantic reasoning network includes: Extract the topological semantic features of high-matching-probability node pairs in cross-domain semantic reasoning networks, and conduct semantic propagation stability analysis of node pairs; Specifically, in this embodiment, the semantic propagation stability of node pair (u,v) is defined as the variance of the matching probability. The specific calculations are as follows: ,in, Let be the matching probability of node pair (u,v) in the k-th iteration. K represents the average matching probability, and K is the number of iterations (e.g., 5 times).

[0062] Furthermore, a variance threshold (e.g., 0.05) is set, and node pairs with variance less than this threshold are identified as node pairs with stable semantic propagation.

[0063] Based on the semantic propagation stability of node pairs, several enterprise-project matching combinations with high semantic stability are generated through a semantic collaborative clustering strategy. Specifically, in this embodiment, for the stable node pairs determined by the aforementioned semantic propagation stability analysis, a semantic collaborative clustering strategy is used to match and combine enterprise nodes and project nodes. The specific implementation of the semantic collaborative clustering strategy is as follows: First, a semantic stability similarity matrix for node pairs is constructed. Based on the semantic propagation stability Var(u,v) of node pair (u,v), the stability variance is transformed into stability similarity (e.g., similarity is defined as...). ); Secondly, collaborative clustering is performed using the similarity matrix. For example, spectral clustering is used for cluster analysis, ensuring that semantic stability similarity meets the clustering requirements (e.g., the default number of clusters in the spectral clustering algorithm is set to a certain value). Node pairs of a class are merged into several semantically stable matching combinations; For example, node pairs with a semantic stability similarity of 0.95 or higher are clustered into the same enterprise-project matching combination to ensure that the semantic propagation characteristics of nodes within each combination are highly consistent.

[0064] Merge matching combinations whose semantic stability exceeds a predetermined threshold to construct a target project information matching set; Specifically, this embodiment defines a semantic stability index for enterprise-project matching combinations. The calculation method is as follows: In the formula, C is the set of matching combinations after clustering; Indicates the number of node pairs in the combination; Let be the variance of semantic propagation stability for node pair (u,v).

[0065] Furthermore, based on the calculations obtained The value is set to a predetermined threshold (e.g., 0.95), and matching combinations that meet or exceed the threshold are merged to construct a target project information matching set.

[0066] For example, if combination C1 contains node pairs ("intelligent agricultural drone R&D enterprise", "smart farmland monitoring project") and ("intelligent agricultural drone R&D enterprise", "unmanned plant protection drone demonstration and promotion project"), and the calculation yields... If the combination C1 meets the predetermined threshold, it will be included in the target project information matching set, thus completing the accurate matching task between enterprises and projects.

[0067] The above embodiments are only used to illustrate the technical methods 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 methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. An automatic matching system for project information of small and medium-sized enterprises, characterized in that, include: The subgraph generation module is used to obtain the initial project information data and domain general knowledge graph of the target enterprise, and generate enterprise feature subgraphs in the domain general knowledge graph based on the domain semantic tags of the initial project information data. The feature acquisition module is used to determine the node diffusion path based on the cross-layer propagation probability of nodes in the enterprise feature subgraph, and generate domain feature transfer vectors in the domain general knowledge graph based on the node diffusion path. The data analysis module is used to construct a multi-view graph inference network based on domain feature transfer vectors, and to determine enterprise-project candidate matching groups based on the node association matrix output by the multi-view graph inference network. The information matching module is used to construct a cross-domain semantic reasoning network based on the semantic reconstruction probability of enterprise-project candidate matching groups in the domain general knowledge graph, and generate a target project information matching set based on the enterprise-project matching probability output by the cross-domain semantic reasoning network.

2. The automatic matching system for project information of small and medium-sized enterprises according to claim 1, characterized in that, The process of generating the enterprise feature sub-graph is as follows: Multimodal semantic features are extracted from the company's initial project information data, and the multimodal semantic features are divided into several semantic subspaces; For each semantic subspace, calculate the cross-semantic distance with nodes in the domain-general knowledge graph, and obtain domain semantic labels through clustering using the semantic distance function; Using domain semantic tags as node activation conditions, and iteratively filtering based on the interaction probability distribution of node activation in the domain general knowledge graph, enterprise feature subgraphs are generated.

3. The automatic matching system for project information of small and medium-sized enterprises according to claim 2, characterized in that, The step of determining the node diffusion path based on the cross-layer propagation probability of nodes in the enterprise feature subgraph includes: Construct a dynamic game model for propagation between nodes in the enterprise feature subgraph, and determine the initial strategy distribution for node propagation based on the dynamic game model; The node diffusion stability state of the initial policy distribution is evaluated using the policy evolution stability condition to obtain the initial cross-layer propagation probability; Under stable propagation conditions, the node diffusion strategy is iteratively updated through the path game benefit function until the node diffusion strategy converges, thus determining the node diffusion path.

4. The automatic matching system for project information of small and medium-sized enterprises according to claim 3, characterized in that, The step of generating domain feature transfer vectors in the domain general knowledge graph based on node diffusion paths includes: The semantic embedding features of each node in the node diffusion path are aggregated through a multi-layer graph attention mechanism to generate a path embedding representation; An embedding feature entropy weight allocation strategy is used to perform unbalanced optimization and adjustment on the semantic contribution of different nodes in the path embedding representation, resulting in an unbalanced optimized path embedding representation. Based on the unbalanced optimized path embedding representation, a domain feature transfer vector with a multi-scale semantic association structure is generated.

5. The automatic matching system for project information of small and medium-sized enterprises according to claim 4, characterized in that, An embedding feature entropy weighting strategy is employed to perform unbalanced optimization and adjustment of the semantic contributions of different nodes in the path embedding representation, including: The initial entropy weight distribution of the embedded features of the nodes in the path embedding representation is computed, and the entropy weight distribution is adjusted with the difference in path propagation mode as a constraint. Based on the inter-node feature entropy weight interaction gain condition, the iterative update rate of the entropy weight distribution is dynamically adjusted until the entropy weight distribution meets the convergence condition. By redistributing the weights of the path embedding representation using the converged entropy weight distribution, an unbalanced optimized path embedding representation is obtained.

6. The automatic matching system for project information of small and medium-sized enterprises according to claim 5, characterized in that, The multi-view graph inference network constructed based on domain feature transfer vectors includes: Multi-scale semantic segmentation is performed on the domain feature transfer vector to generate several sub-vectors with semantic granularity differences; Based on the matching strength between the sub-vector granularity difference and the domain-general knowledge graph, a graph structure adaptive learning strategy is adopted to construct a multi-view initial sub-network; By leveraging the advantages of cross-granularity semantic propagation between subnetworks, the initial subnetworks are merged to construct the final multi-view graph reasoning network.

7. The automatic matching system for project information of small and medium-sized enterprises according to claim 6, characterized in that, The process of determining enterprise-project candidate matching groups based on the node association matrix output by the multi-view graph inference network includes: The propagation features of the node association matrix are mapped to a low-dimensional semantic space through probabilistic manifold learning to generate a semantically compact representation; By leveraging the semantic similarity of nodes and frequent patterns of topological propagation in compact representations, several initial enterprise-project matching structures are formed. Based on the semantic and topological coupling degree of the initial matching structure, candidate matching groups of enterprises and projects with stable structural coupling are selected.

8. The automatic matching system for project information of small and medium-sized enterprises according to claim 7, characterized in that, The construction of a cross-domain semantic reasoning network based on the semantic reconstruction probability of enterprise-project candidate matching groups in the domain general knowledge graph includes: Calculate the heterogeneous mapping relationship between the topological propagation structure and the semantic structure of the domain-general knowledge graph among candidate matching group node pairs, and identify key semantic reconstruction nodes; Based on the cross-domain semantic reconstruction strength of key semantic reconstruction nodes, the initial cross-domain semantic propagation path is dynamically optimized and adjusted; Based on the propagation entropy of the optimized semantic propagation path, a cross-domain semantic reasoning network is constructed.

9. The automatic matching system for project information of small and medium-sized enterprises according to claim 8, characterized in that, The process of generating a target project information matching set based on the enterprise-project matching probability output by the cross-domain semantic reasoning network includes: Extract the topological semantic features of high-matching-probability node pairs in cross-domain semantic reasoning networks, and conduct semantic propagation stability analysis of node pairs; Based on the semantic propagation stability of node pairs, several enterprise-project matching combinations with high semantic stability are generated through a semantic collaborative clustering strategy. Matches with semantic stability exceeding a predetermined threshold are merged to construct a target project information matching set.

Images