Enterprise resource recommendation method and device based on scene graph, medium and equipment

Abstract

This application provides a method, apparatus, medium, and device for enterprise resource recommendation based on scene graphs, belonging to the field of data processing technology. The method includes: acquiring enterprise resource demand information and identifying demand keywords; matching basic scenes from a scene graph library based on the demand keywords; expanding keywords according to a weight matrix; matching extended scenes from the scene graph library based on the extended keywords, and forming a mapping scene based on the extended scene and the basic scene; extracting a set of candidate resources corresponding to the resource demand information from the mapping scene; constructing a compatibility matrix between candidate resources, a matching matrix between candidate resources and the mapping scene, and a cost matrix between existing resources and candidate resources based on the enterprise's existing resources, the mapping scene, and the candidate resources in the candidate resource set; and calculating recommended resources to be pushed to the enterprise based on these three matrices. This application can improve the accuracy and practicality of enterprise resource recommendation.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus, medium and device for enterprise resource recommendation based on scene graphs. Background Technology

[0002] The challenges of digital transformation in manufacturing are characterized by their complexity and depth: On the surface, the transformation needs faced by enterprises appear scattered, with clearly defined pain points and seemingly independent phenomena (such as long changeover times on assembly lines, excessive energy consumption across multiple plants, and delayed equipment failure warnings). However, deeper analysis reveals that these needs are intertwined with collaborative issues across multiple business segments, including R&D, manufacturing, operation and maintenance, business management, and supply chain management. This systemic complexity, where a change in one area can have far-reaching consequences, makes exploring transformation paths far more difficult than solving a single problem. Existing technologies struggle to provide effective solutions, primarily due to the following:

[0003] First, the descriptions of enterprise transformation needs are vague, and the mapping between needs and scenarios is inaccurate. Manufacturing enterprises often lack professional knowledge of digital scenarios and can only describe their needs through business phenomena such as high production costs and low production efficiency, failing to clearly define specific needs and potential needs (e.g., only mentioning reducing energy consumption in multiple plants, but not mentioning excessive energy consumption at an air compressor station in a certain plant or the collaborative utilization of waste heat in multiple plants). Existing technologies mostly rely on single semantic similarity search to match needs and scenarios, without considering the hierarchical correlation of scenarios (e.g., "energy consumption optimization" needs to be associated with a multi-layered structure of "production manufacturing - energy management - multi-plant collaboration") and the collaborative relationships between scenarios. This results in mapping results that only cover the surface scenarios, leading to large deviations in demand positioning and high transformation communication costs.

[0004] Secondly, digital resource recommendations are fragmented, and the connections between elements and resources are disconnected. Manufacturing transformation requires the coordinated support of four types of elements: data elements, knowledge models, software tools, and human skills. Furthermore, elements and resources are interdependent (e.g., optimizing a digital twin assembly line requires supporting modeling tools, mechanism models, assembly data, and modeling talent). However, existing technologies often adopt an isolated recommendation model of "single element - single resource," recommending only a certain type of resource (e.g., only recommending energy consumption monitoring tools without linking analysis models and data solutions), or failing to consider resource compatibility. Enterprises must then manually combine numerous complex resources, resulting in low resource utilization and difficulties in implementing transformation.

[0005] Finally, industry scenario data lacks structured integration and is not adapted to the differentiated transformation foundations of enterprises. Manufacturing scenarios encompass a multi-level structure of "industry → business process → scenario → value tag," and are closely linked across business scenarios (e.g., flexible assembly on the final assembly line requires linkage with component manufacturing and supply chain material scheduling). However, existing scenario data is mostly unstructured text or scattered storage, making it impossible to quantify scenario relationships and support cross-business collaborative transformation. Furthermore, existing technology-based resource recommendations do not consider existing enterprise resources (e.g., large enterprises have deployed ERP systems, while SMEs still use manual ledgers), easily leading to resource incompatibility, duplicate recommendations, or excessively high upgrade costs, failing to meet the differentiated needs of enterprises.

[0006] In summary, existing technologies cannot solve the core problems of inaccurate mapping between demand and scenario, fragmented resource recommendations, and insufficient adaptability during the transformation of manufacturing enterprises. There is an urgent need for a technical solution that integrates structured scenario data and accurately matches demand and resources to reduce the transformation costs of enterprises and promote the efficient implementation of demand. Summary of the Invention

[0007] The purpose of this application is to provide a method, apparatus, medium, and device for enterprise resource recommendation based on scene graphs, so as to solve at least one of the above-mentioned technical problems.

[0008] In a first aspect, this application provides a method for recommending enterprise resources based on scene graphs, the method comprising:

[0009] Obtain enterprise resource demand information and identify demand keywords from the resource demand information;

[0010] Based on the aforementioned requirement keywords, a first number of basic scenarios are matched from the scenario graph library;

[0011] Based on the weight matrix of keywords preset in the scene graph of the basic scene, the keywords are expanded to form extended keywords corresponding to the required keywords;

[0012] Based on the extended keywords, a second number of extended scenarios are matched from the scenario graph library, and a mapping scenario matching the resource requirement information is formed according to the extended scenarios and the basic scenarios.

[0013] Extract a set of candidate resources corresponding to the resource demand information from the mapped scenario;

[0014] Based on the enterprise's existing resources, mapping scenarios, and candidate resources in the candidate resource set, construct the compatibility matrix between candidate resources, the matching matrix between candidate resources and mapping scenarios, and the cost matrix between existing resources and candidate resources;

[0015] Based on the compatibility matrix, matching matrix, and cost matrix, recommended resources are calculated for pushing to the enterprise.

[0016] Optionally, the step of matching a first number of basic scenarios from the scene graph library based on the demand keywords includes: converting the demand keywords into a first embedding vector, and performing a first scene matching from the scene graph library based on the first embedding vector to obtain a first number of basic scenarios.

[0017] Optionally, the step of matching a second number of extended scenes from the scene graph library based on the extended keywords includes: converting the extended keywords into a second embedding vector, and performing a second scene matching from the scene graph library based on the second embedding vector to obtain a second number of extended scenes.

[0018] Optionally, the step of performing a first scene matching from the scene graph library based on the first embedding vector to obtain a first number of basic scenes includes:

[0019] Obtain the text block subsets associated with each scene from the scene graph library, and calculate the average similarity based on the first embedding vector and the embedding vector set of the associated text block subsets of each scene;

[0020] The scenes are sorted from high to low according to their average similarity, and the scene with the highest similarity among them is selected as the base scene.

[0021] Optionally, the step of constructing a compatibility matrix between candidate resources, a matching matrix between candidate resources and mapping scenarios, and a cost matrix between existing resources and candidate resources based on the enterprise's existing resources, mapping scenarios, and candidate resources in the candidate resource set includes:

[0022] Extract the first keyword of the candidate resources, the second keyword of the mapping scenario, and the third keyword of the existing resources;

[0023] In the mapping scenario, find the corresponding element values ​​between the first keyword, the second keyword, and the third keyword in the preset keyword weight matrix;

[0024] Based on the found element values, the compatibility matrix, matching matrix, and cost matrix are constructed respectively.

[0025] Optionally, the element values ​​in the compatibility matrix are positively correlated with the corresponding found element values; the element values ​​in the matching matrix are positively correlated with the corresponding found element values; and the element values ​​in the cost matrix are negatively correlated with the corresponding found element values.

[0026] Optionally, the compatibility matrix element values ​​in It is calculated in the following way:

[0027] Acquisition and candidate resources Associated text block subsets and candidate resources For related text block subsets, calculate the first average similarity between two text block subsets. ;

[0028] Based on the representation in the weight matrix and element value and Calculate and The compatibility between them, as a compatibility matrix element values ​​in .

[0029] Optionally, the matching matrix element values ​​in It is calculated in the following way:

[0030] Acquisition and candidate resources Associated text block subsets and mapping scenarios For related text block subsets, calculate the second average similarity between two text block subsets. ;

[0031] Based on the representation in the weight matrix and element value and Calculate and The matching degree between them is used as the matching matrix. element values ​​in .

[0032] Optionally, the cost matrix element values ​​in It is calculated in the following way:

[0033] Acquisition and candidate resources Associated text block subsets and existing resources For related text block subsets, calculate the third average similarity between two text block subsets. ;

[0034] Based on the representation in the weight matrix and element value and Calculate and The cost matrix between them element values ​​in .

[0035] Optionally, before constructing the compatibility matrix between candidate resources, the matching matrix between candidate resources and mapping scenarios, and the cost matrix between existing resources and candidate resources based on the enterprise's existing resources, mapping scenarios, and candidate resources in the candidate resource set, the method further includes:

[0036] Acquire the enterprise's digital asset data, extract keywords from the digital asset data, and obtain existing resource keywords;

[0037] The existing resource keywords are matched with a preset resource classification dictionary to determine the type and attribute parameters of the existing resources, thus completing the identification of existing resources.

[0038] A second aspect of this application provides an enterprise resource recommendation device based on a scene graph, the device comprising:

[0039] The demand information acquisition module is used to acquire the enterprise's resource demand information and identify demand keywords from the resource demand information;

[0040] The scene recognition module is used to match a first number of basic scenes from the scene graph library based on the demand keywords; expand the keywords according to the weight matrix of the keywords of the basic scenes in the scene graph to form extended keywords corresponding to the demand keywords; match a second number of extended scenes from the scene graph library based on the extended keywords; and form a mapping scene that matches the resource demand information according to the extended scenes and the basic scenes.

[0041] The resource recommendation module is used to extract a set of candidate resources corresponding to the resource demand information from the mapping scenario; construct a compatibility matrix between candidate resources, a matching matrix between candidate resources and the mapping scenario, and a cost matrix between existing resources and candidate resources based on the enterprise's existing resources, the mapping scenario, and the candidate resources in the candidate resource set; and calculate recommended resources to be pushed to the enterprise based on the compatibility matrix, the matching matrix, and the cost matrix.

[0042] In a third aspect, this application provides a computer-readable storage medium storing executable instructions that, when executed by a processor, cause the processor to perform the method described in any embodiment of this application.

[0043] In a fourth aspect, this application provides an electronic device comprising: one or more processors; and a memory for storing one or more programs, which, when executed by the one or more processors, cause the one or more processors to perform the method as described in any embodiment of this application.

[0044] The enterprise resource recommendation method, apparatus, medium, and equipment based on scenario graphs in this application first extract basic scenarios based on identified demand keywords. Then, the demand keywords are expanded based on the basic scenarios to obtain extended keywords. Based on these extended keywords, scenario identification is performed again to obtain extended scenarios. Finally, the mapped scenarios corresponding to the enterprise resource demands are obtained by combining these extended scenarios with the demand scenarios. Through hierarchical matching of "demand keywords → basic scenarios → extended scenarios → mapped scenarios," vague enterprise demands (such as reducing final assembly costs) are accurately mapped to standardized business scenarios (such as multi-plant management - cost control) in the scenario graph, ensuring that recommended resources are highly consistent with the actual business links of the enterprise, and improving the accuracy and comprehensiveness of the identified mapped scenarios. After obtaining the mapping scenario, appropriate candidate resources are extracted from the mapping scenario. The compatibility matrix, matching matrix, and cost matrix are constructed by combining the existing resources, the mapping scenario, and the candidate resources. The comprehensive score of the candidate resources is calculated based on these three matrices. Based on the comprehensive score, an appropriate number of candidate resources are selected as recommended resources. This realizes the quantification of the relationship between resources and enterprise needs, replaces the traditional manual recommendation that relies on expert experience, reduces subjective bias, makes resource recommendations more in line with scenario needs, and improves the accuracy and practicality of resource recommendations. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0046] Figure 1 This is a flowchart illustrating an enterprise resource recommendation method based on scene graphs in one embodiment;

[0047] Figure 2 This is a flowchart illustrating the process of constructing a compatibility matrix between candidate resources, a matching matrix between candidate resources and mapping scenarios, and a cost matrix between existing resources and candidate resources based on the enterprise's existing resources, mapping scenarios, and candidate resources in the candidate resource set, in one embodiment.

[0048] Figure 3 This is a schematic diagram of the structure of an enterprise resource recommendation device based on scene graph in one embodiment;

[0049] Figure 4 This is a schematic diagram of the structure of an enterprise resource recommendation device based on scene graphs in another embodiment;

[0050] Figure 5 This is a schematic diagram of the structure of an electronic device in one embodiment. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0052] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0053] For example, the terms "first," "second," etc., used in this application may be used herein to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from another element.

[0054] For example, the terms "comprising" or "including" used in this application indicate the presence of features, steps, operations and / or components, but do not exclude the presence or addition of one or more other features, steps, operations or components.

[0055] This application provides a method for enterprise resource recommendation based on scene graphs, combined with Figure 1 As shown, the method includes:

[0056] Step 110: Obtain the enterprise's resource demand information and identify demand keywords from the resource demand information.

[0057] In this embodiment, resource requirement information refers to the demands made by enterprises related to their business optimization goals. Specifically, it can be the demands made by enterprises during their digital transformation process related to their business optimization goals, covering dimensions such as cost, efficiency, and quality. For example, an engineering machinery company might propose "reducing the changeover time of the final assembly line to within 30 minutes and reducing the monthly cost of the multi-plant final assembly process by 5%." Resource requirement information can be a combination of one or more types of information, such as text, voice, video, and images. Requirement keywords represent terms extracted from the resource requirement information that characterize the core demands.

[0058] Electronic devices can collect enterprise (transformation) resource demand information through a pre-set large language model (or through a BERT pre-trained semantic model or other related models), and process the (transformation) resource demand information to remove redundant information and extract corresponding demand keywords (such as final assembly line, changeover time, cost reduction, etc.).

[0059] Through a pre-set large language model, enterprises can be guided to input the required (transformation) resource demand information. In addition to key demand information, this (transformation) resource demand information can also include basic enterprise information, such as the enterprise's industry, the business process and business activities corresponding to the transformation demand, and additional information such as the enterprise's current existing resources and difficulties encountered in the transformation process.

[0060] Specifically, based on a pre-trained large language model (such as the DeepSeek model) and a corresponding vector embedding model (such as the DeepSeek-embed model), the required text can be segmented into multiple text blocks, each containing a certain number of tokens, for example, ≤512 tokens. Each segmented text block is converted into an embedding vector, forming an embedding vector set. This embedding vector set is then clustered using a pre-defined clustering method to obtain multiple topic groups containing the embedding vectors from this set. A pre-defined number of text block samples are selected for each topic group, and then combined with historical resource keyword samples (such as assembly efficiency, factory cost, etc.) to construct a prompt.

[0061] The text block samples can be selected by choosing the n1 samples closest to the cluster center, or by randomly selecting n2 samples. For example, based on the text block samples "The current changeover time of the excavator assembly line in the East China plant is about 60 minutes..." and "The changeover time needs to be reduced to within 30 minutes...", combined with historical keywords "assembly efficiency" and "plant cost", a preset large language model is used to extract and refine 5 keywords with a length ≤ 10 tokens, ultimately obtaining a set of demand keywords. For example, the extracted set of demand keywords could be {assembly line changeover time, multiple plants, assembly cost, 30 minutes, 5%}. n1 and n2 can be any suitable positive integers.

[0062] Step 120: Based on the required keywords, match the first number of basic scenarios from the scenario graph library.

[0063] In this embodiment, the scenario graph library contains various scenario graphs. A scenario graph is a systematic graph formed by combining several scenarios according to industry business logic and discourse, representing a standardized, structured, and modular expression of the industry. Relying on industry digital knowledge, the scenario graph is a structured graph that generates a standardized representation of the entire chain of industry digital transformation knowledge through end-to-end processes. One or more scenario graphs can be constructed for different industries and enterprise business activities.

[0064] Each scenario graph contains a structured architecture including multi-level nodes, edges between nodes, and weight matrices. Multi-level nodes can include industry link nodes (such as the upstream and downstream links in the construction machinery industry, such as "upstream raw material procurement → midstream assembly and manufacturing → downstream operation and maintenance services"), business activity nodes (such as R&D design, production and manufacturing, operation and maintenance services, business management, and supply chain management), digital scenario nodes (including main scenarios and / or sub-scenarios, such as specific scenarios under the business activity of "production and manufacturing → final assembly and integration → flexible assembly"), digital element nodes (such as elements like software tools, knowledge models, data elements, and talent skills), and value tag nodes (such as quantifiable scenario values ​​like cost reduction and efficiency improvement). Edges between nodes can include belonging edges and supporting edges, used to connect directed edges between nodes of different levels, reflecting the "belonging / supporting" relationship. Belonging edge: Industry segment → Business activity (e.g., "Midstream assembly manufacturing → Production manufacturing"), Business activity → Digital scenario (e.g., "Production manufacturing → Flexible assembly"); Supporting edge: Digital element → Digital scenario (e.g., "AGV intelligent scheduling software → Flexible assembly"), Digital scenario → Value tag (e.g., "Flexible assembly → Efficiency improvement").

[0065] The weight matrix quantifies the strength of associations between nodes. Each node can be represented by a keyword, and the weight matrix is ​​thus the weight matrix between these keywords. Each element in the weight matrix (i.e., the weight value between two keywords) represents the weight (degree of association) between the corresponding two associated edges (keywords). The stronger the association between two keywords, the larger their corresponding element value (weight value). This weight can include the weight of the belonging edge and the weight of the supporting edge. The weight of the belonging edge is calculated based on "industry-preset knowledge rules + vertical category large model association strength" (e.g., the weight between "final assembly manufacturing - production manufacturing" is 0.92); the weight of the supporting edge is calculated based on "node label semantic similarity + size of the intersection of associated text blocks" (e.g., the weight between "AGV software - flexible assembly" is 0.85). The matrix elements in the weight matrix are non-negative and symmetric.

[0066] The digital scene nodes of the scene graph include main scenes and / or sub-scenes (such as the main scene of flexible assembly and the sub-scenes of AGV scheduling optimization and changeover process collaboration). The basic scenes refer to the scene nodes related to the demand keywords selected from the scene graph library. They are a subset of highly relevant scenes that are directly matched with the enterprise's needs and serve as the hub for the implementation of the scene graph and the enterprise's needs.

[0067] The basic scenario can be determined based on the semantic similarity of the requirement keyword set with the tags in the scenario node and the matching degree of related data. The matched basic scenario can include one or more. For example, if a company's requirement is to shorten the final assembly changeover time, the matched basic scenario could be the flexible assembly scenario in a scenario map. The first quantity can be any preset and suitable number, such as 3.

[0068] Each scene graph in the scene graph library contains scene keywords for each corresponding scene. Electronic devices can match the scene keywords in each scene graph based on the set of demand keywords. This matching can include precise matching and semantic matching to determine one or more of the most matching scene graphs from the scene graph library, and further refine the matching scene graphs to match one or more specific scenes as the corresponding base scenes.

[0069] Step 130: Expand keywords based on the weight matrix of keywords preset in the scene graph of the basic scene to form extended keywords corresponding to the required keywords.

[0070] In this embodiment, the extended keywords are keywords selected from other keywords based on the relationship between the element values ​​of the demand keyword and other keywords in the corresponding weight matrix. This selection method can involve randomly selecting a preset number of keywords from the corresponding weight matrix. For example, it can involve randomly selecting from a set of keywords in the weight matrix whose weight values ​​with the demand keyword exceed a preset threshold, and extracting a preset number of keywords as extended keywords.

[0071] Alternatively, keywords with weight values ​​exceeding a preset threshold that are selected from the weight matrix can be used to form a candidate keyword set. Then, from the historical keyword identification records, the current demand keyword is used as the target keyword. The number of times each candidate keyword in the candidate keyword set is used as a demand keyword or extended keyword in the historical records is used as a demand keyword or extended keyword simultaneously forms a ranking of the number of times each candidate keyword is used. The candidate keyword with the highest ranking and a preset number of times is selected as an extended keyword of the target keyword.

[0072] Understandably, extended keywords are intended to supplement inaccurate or incomplete descriptions of resource requirements by enterprises, rather than simply semantically expanding upon existing requirement keywords. The requirement keywords identified in step 110 can include multiple keywords, and each requirement keyword can correspond to one or more extended keywords. Therefore, an extended keyword for one requirement keyword may be another requirement keyword identified in step 110. That is, there is an intersection between the set of requirement keywords and the set of extended keywords. By selecting keywords from the weight matrix, the accuracy of extended keyword selection can be improved, thus enhancing the comprehensiveness of keyword identification for resource recommendations.

[0073] Step 140: Based on the extended keywords, a second number of extended scenarios are matched from the scenario graph library, and a mapping scenario matching the resource requirement information is formed according to the extended scenarios and the basic scenarios.

[0074] In this embodiment, the identified extended keywords can be merged with the basic keyword set. Using the merged keywords as a basis, scene matching is then performed again from the scene graph library. This matching process is the same as step 120, and the matched scene is the extended scene, which will not be described further here. The second quantity can also be any preset, suitable quantity, such as 3 or 2.

[0075] Understandably, the matched extended scenes may overlap with the base scenes. Electronic devices can take the union of the base scene set and the extended scene set, using all scenes in the union as mapping scenes, or select a preset number of scenes with the highest matching degree in the union as mapping scenes.

[0076] Step 150: Extract a set of candidate resources corresponding to the resource demand information from the mapped scenario.

[0077] In this embodiment, the resources in each scenario may include various resources required by the enterprise in the corresponding scenario, such as software tools, knowledge models, data elements, talent skills and other types of resources required for enterprise transformation. Each type of resource may include multiple resources.

[0078] Based on the obtained mapping scenario, a preset number of resources that best match the resource demand information under the mapping scenario can be extracted as a candidate resource set, or all resources under the mapping scenario can be used as candidate resources.

[0079] In one embodiment, a suitable number of candidate resources can be selected from all resources in the mapped scenario using a pre-defined principal component analysis method, forming a candidate resource set.

[0080] Specifically, feature vectors are extracted from all resources in all mapping scenarios; a resource feature matrix is ​​constructed based on these feature vectors. The feature vectors include multiple feature dimensions such as keyword matching degree between resources and transformation demand information, association weight between resources and mapping scenarios, historical application effect scores of resources, and adaptation cost coefficients of resources. Each feature dimension, after standardization, takes values ​​in the range [0,1]. Let the number of resources in all mapping scenarios be m1, and the feature vector dimensions be n4. Then the feature matrix X is an m1×n4 matrix, where x(i,j) represents the j-th feature value of the i-th resource.

[0081] Calculate the covariance matrix C of the feature matrix X, and obtain the eigenvalues ​​and corresponding eigenvectors of the covariance matrix C through eigenvalue decomposition; determine the number of principal components k: select the smallest k value whose cumulative variance contribution rate exceeds the preset threshold; calculate the principal component score of each resource, and select the top n3 resources in terms of score as the candidate resource set, where n3 is the preset threshold for the number of candidate resources.

[0082] Principal component analysis can be used to screen candidate resources objectively, thus avoiding the subjectivity of manual screening. Furthermore, the variance contribution rate can be used to quantify the screening criteria, thereby improving the reliability of the candidate resource set.

[0083] Step 160: Based on the enterprise's existing resources, mapping scenarios, and candidate resources in the candidate resource set, construct the compatibility matrix between candidate resources, the matching matrix between candidate resources and mapping scenarios, and the cost matrix between existing resources and candidate resources.

[0084] In this embodiment, the compatibility matrix is ​​an m×m matrix (m is the number of candidate resources) that characterizes the collaborative compatibility between candidate resources. The element values ​​in the compatibility matrix represent the compatibility between two corresponding resources. The compatibility can be calculated by the intersection ratio and average similarity of the resource-related text block subsets. The larger the value, the stronger the compatibility.

[0085] The matching matrix is ​​an m×n matrix representing the fit between candidate resources and the mapped scene (m is the number of candidate resources and n is the number of mapped scenes). The element values ​​in the matching matrix represent the matching degree between the corresponding candidate resource and the mapped scene. It can also be calculated by the average similarity between the resource text block subset and the scene text block subset. The larger the value, the higher the fit.

[0086] The cost matrix represents the p×m matrix of the adaptation cost between existing resources and candidate resources (p is the number of existing resources and m is the number of candidate resources). The element values ​​can also be calculated by the average similarity between the resource text block subset and the scene text block subset. The smaller the value, the lower the adaptation cost.

[0087] Step 170: Calculate the recommended resources to be pushed to the enterprise based on the compatibility matrix, matching matrix, and cost matrix.

[0088] Based on these three matrices, a comprehensive score is calculated for each candidate resource, and a predetermined number of candidate resources with the highest scores are selected as the enterprise's recommended resources. This comprehensive score can be obtained through operations between the three matrices. For example, the compatibility matrix and the matching matrix can be multiplied to obtain a first matrix. The values ​​of elements in this first matrix representing the same candidate resource can be averaged to obtain a first average. The values ​​of elements in the cost matrix representing the same candidate resource can also be averaged to obtain a second average. Finally, the first average and the second average are weighted and summed to obtain the comprehensive score for the corresponding candidate resource. The weights between the first and second averages can be set according to requirements.

[0089] The enterprise resource recommendation method based on scenario graphs in this application first extracts basic scenarios based on identified demand keywords. Then, it expands the demand keywords based on the basic scenarios to obtain extended keywords. Based on these extended keywords, it performs scenario identification again to obtain extended scenarios. Finally, it combines these extended scenarios with the demand scenarios to obtain the mapping scenario corresponding to the enterprise resource demand. Through hierarchical matching of "demand keywords → basic scenarios → extended scenarios → mapping scenarios," it accurately maps vague enterprise demands (such as reducing final assembly costs) to standardized business scenarios (such as multi-plant management - cost control) in the scenario graph, ensuring that recommended resources are highly consistent with the enterprise's actual business links and improving the accuracy and comprehensiveness of the mapping scenario identification.

[0090] After obtaining the mapping scenario, appropriate candidate resources are extracted from the mapping scenario. The compatibility matrix, matching matrix, and cost matrix are constructed by combining the existing resources, the mapping scenario, and the candidate resources. The comprehensive score of the candidate resources is calculated based on these three matrices. Based on the comprehensive score, an appropriate number of candidate resources are selected as recommended resources. This realizes the quantification of the relationship between resources and enterprise needs, replaces the traditional manual recommendation that relies on expert experience, reduces subjective bias, makes resource recommendations more in line with scenario needs, and improves the accuracy and practicality of candidate resource recommendations.

[0091] In one embodiment, matching a first number of basic scenarios from the scene graph library based on the demand keywords includes: converting the demand keywords into a first embedding vector, and performing a first scene matching from the scene graph library based on the first embedding vector to obtain a first number of basic scenarios; matching a second number of extended scenarios from the scene graph library based on the extended keywords includes: converting the extended keywords into a second embedding vector, and performing a second scene matching from the scene graph library based on the second embedding vector to obtain a second number of extended scenarios.

[0092] In this embodiment, a pre-trained BERT-base model can be invoked to encode each keyword, outputting an initial vector corresponding to each model. The dimension of this initial vector can be a pre-set appropriate dimension, such as 768 dimensions. The initial vectors for each required keyword are then weighted and summed, and the resulting vector is used as the first embedding vector. This weighted summation can be achieved by averaging the values ​​of each dimension.

[0093] After obtaining the first embedding vector, the similarity between the first embedding vector and the embedding vector of each scene node in the scene graph is calculated. After obtaining the similarity with all scene nodes, the first number of scenes with the highest similarity are selected as the base scenes.

[0094] This similarity can be cosine similarity, or other calculation methods. (Embedded vector) and embedding vector The formula for calculating the cosine similarity between them is: .

[0095] Optional, embedding vector and embedding vector The similarity between them can also be calculated using the following similarity function formula: .

[0096] and This represents a local normalization constant calculated separately for each node / keyword. This similarity function smoothly maps the keywords / nodes corresponding to two embedding vectors to a similarity weight value between 0 and 1.

[0097] Similarly, the matching method for the second scene is the same as that for the first scene. A pre-trained BERT-base model can be invoked to encode each keyword containing both the demand keyword and the extended keywords, outputting an initial vector for each model. These initial vectors are then weighted and summed to obtain the second embedding vector. The similarity between this second embedding vector and the embedding vector of each scene node in the scene graph is calculated in the same way, and the extended scene is determined based on this similarity.

[0098] By using embedded vectors for scene matching, the accuracy of scene matching can be improved.

[0099] In one embodiment, the step of performing first scene matching from the scene graph library based on the first embedding vector to obtain a first number of basic scenes includes: obtaining a subset of text blocks associated with each scene from the scene graph library; calculating the average similarity based on the first embedding vector and the embedding vector set of the associated text block subsets of each scene; sorting each scene from high to low according to the average similarity, and selecting the first number of scenes before sorting as basic scenes.

[0100] In this embodiment, the subset of text blocks in the scene is a set of topic group text blocks selected by cyclically clustering the text block set of the industry digital knowledge base using K-means clustering and spectral clustering. The selected topic group text blocks include the n1 text blocks closest to the cluster center and the n2 text blocks randomly selected.

[0101] To calculate the average similarity, we can first calculate the similarity between the first embedding vector and the embedding vector of each text block within the text block subset, and then take the average of the calculated similarities to obtain the average similarity between the scene and the requirement keywords. This similarity can be calculated using the cosine similarity formula or similarity function mentioned above.

[0102] In one embodiment, such as Figure 2 As shown, the construction of a compatibility matrix between candidate resources, a matching matrix between candidate resources and mapping scenarios, and a cost matrix between existing resources and candidate resources, based on the enterprise's existing resources, mapping scenarios, and candidate resources in the candidate resource set, includes:

[0103] Step 210: Extract the first keyword of the candidate resource, the second keyword of the mapping scenario, and the third keyword of the existing resource.

[0104] Step 220: Find the corresponding element values ​​between the first keyword, the second keyword, and the third keyword in the preset keyword weight matrix in the mapping scenario.

[0105] Step 230: Construct the compatibility matrix, matching matrix, and cost matrix based on the found element values.

[0106] In this embodiment, for each candidate resource, keywords describing the candidate resource are extracted from its corresponding mapping scenario and used as the first keyword; for the mapping scenario, keywords describing the mapping scenario are also obtained as the second keyword; for existing resources, their corresponding descriptive information is also obtained to form the third keyword. It is understood that the number of first, second, and third keywords can all be multiple.

[0107] For the extracted first, second, and third keywords, find the element values ​​representing the weights among the first, second, and third keywords in the weight matrix of the scene graph where the corresponding mapping scene is located. Based on the found element values, construct the aforementioned compatibility matrix, matching matrix, and cost matrix.

[0108] Optionally, the element values ​​in the compatibility matrix are positively correlated with the corresponding element values ​​found; the element values ​​in the matching matrix are positively correlated with the corresponding element values ​​found; and the element values ​​in the cost matrix are negatively correlated with the corresponding element values ​​found.

[0109] For example, the element values ​​in the found weight matrix can be directly used as the element values ​​of the corresponding two candidate resources in the compatibility matrix; the element values ​​in the found weight matrix can be directly used as the element values ​​of the corresponding candidate resources and mapping scenarios in the matching matrix; the element values ​​in the found weight matrix can be subtracted from the element value by 1 and used as the element values ​​of the corresponding candidate resources and existing resources in the cost matrix.

[0110] In one embodiment, the compatibility matrix element values ​​in The following calculation method was used to obtain candidate resources. Associated text block subsets and candidate resources For related text block subsets, calculate the first average similarity between two text block subsets. Based on the representation in the weight matrix and element value and Calculate and The compatibility between them, as a compatibility matrix element values ​​in .

[0111] The matching matrix element values ​​in The following calculation method was used to obtain candidate resources. Associated text block subsets and mapping scenarios For related text block subsets, calculate the second average similarity between two text block subsets. Based on the representation in the weight matrix and element value and Calculate and The matching degree between them is used as the matching matrix. element values ​​in .

[0112] The cost matrix element values ​​in The following calculation method was used to obtain candidate resources. Associated text block subsets and existing resources For related text block subsets, calculate the third average similarity between two text block subsets. Based on the representation in the weight matrix and element value and Calculate and The cost matrix between them element values ​​in .

[0113] In this embodiment, the first average similarity, the second average similarity, and the third average similarity can all be calculated using the cosine similarity formula or similarity calculation function described above.

[0114] Specifically, ;

[0115] ;

[0116] .

[0117] In one embodiment, before constructing the compatibility matrix between candidate resources, the matching matrix between candidate resources and the mapping scenario, and the cost matrix between existing resources and candidate resources based on the enterprise's existing resources, mapping scenarios, and candidate resources in the candidate resource set, the method further includes: acquiring the enterprise's digital asset data, extracting keywords from the digital asset data to obtain existing resource keywords; matching the existing resource keywords with a preset resource classification dictionary to determine the type and attribute parameters of the existing resources, and completing the identification of existing resources.

[0118] In this embodiment, the enterprise's digital asset data can be obtained from document content uploaded by the enterprise and text information provided in the dialog window. The obtained digital asset data can also be parsed using a large language model to identify keywords (existing resource keywords) corresponding to resources such as the aforementioned data elements, knowledge models, software tools, and talent skills.

[0119] Electronic devices pre-build a standardized resource classification dictionary, which includes type definitions, feature keywords, and required / optional attribute parameters. The extracted existing resource keywords are semantically matched with the feature keywords in the classification dictionary to obtain the existing resources corresponding to the existing resource keywords, thus completing the identification of existing resources.

[0120] In one embodiment, the above method further includes a scene atlas generation process, which includes:

[0121] S1. Keyword Extraction: Collect a set of text blocks T={t1,t2,...,tn} from the digital knowledge base of the industry corresponding to the scene graph to be constructed. Convert each text block ti into a corresponding embedding vector v(ti) using a vector embedding model, thus forming an embedding vector set V={v(t1),v(t2),...,v(tn)}. Use K-means clustering and spectral clustering to iteratively cluster the embedding vector set V into C1 topic groups. For each topic group, select a preset number of text blocks and historical keyword samples to construct a prompt. Extract keywords using a pre-trained large language model. After filtering and refining, obtain a keyword set K. The keyword set K includes keywords corresponding to industry nodes, business activity nodes, scene nodes, resource nodes, pain point nodes, and value tag nodes.

[0122] S2. Text Block Graph Structure Construction: Constructing the text block graph structure Where T is a text block node, Let w(v(ti),v(tj)) be the edge weight matrix. Calculate the similarity w(v(ti),v(tj)) between the embedding vectors of any two text blocks. This similarity can also be calculated using the cosine similarity formula or similarity function described above. Based on the calculated similarities, only the similarities of the m1 nearest neighbors of each node (e.g., m1 takes values ​​from 20 to 30) are retained to construct a sparse matrix. Then, through symmetry processing, the final sparse and symmetric weight matrix is ​​obtained. ;

[0123] S3. Keyword Map Construction: Constructing a keyword map Where K is the keyword node; for each keyword Select p1 positive sample text blocks and p2 negative sample text blocks, and use Laplace semi-supervised learning to obtain their associated text block subsets. ; Calculate any two keywords and Size of the intersection of subsets of related text blocks After obtaining the intersection size among all keywords, it is normalized and used as the corresponding edge weight, so that the weight of the i-th row and j-th column in the matrix is... The weight in the j-th row and i-th column Equal, that is .in, Keywords relative to keywords The weight, Keywords relative to keywords The weight.

[0124] S4. Integrating Keyword Maps The nodes, edges, and weight matrix form a scenario graph containing hierarchical relationships of "industry → business activities → scenario → resources → pain points → value tags".

[0125] In one embodiment, such as Figure 3 As shown, an enterprise resource recommendation device based on scene graph is provided, the device comprising:

[0126] The demand information acquisition module 310 is used to acquire the enterprise's resource demand information and identify demand keywords from the resource demand information.

[0127] The scene recognition module 320 is used to match a first number of basic scenes from the scene graph library based on the demand keywords; expand the keywords according to the weight matrix of the keywords preset in the scene graph where the basic scenes are located, to form extended keywords corresponding to the demand keywords; match a second number of extended scenes from the scene graph library based on the extended keywords, and form a mapping scene that matches the resource demand information according to the extended scenes and the basic scenes.

[0128] The resource recommendation module 330 is used to extract a set of candidate resources corresponding to the resource demand information from the mapping scenario; construct a compatibility matrix between candidate resources, a matching matrix between candidate resources and the mapping scenario, and a cost matrix between existing resources and candidate resources based on the enterprise's existing resources, the mapping scenario, and the candidate resources in the candidate resource set; and calculate recommended resources to be pushed to the enterprise based on the compatibility matrix, the matching matrix, and the cost matrix.

[0129] In one embodiment, the scene recognition module 320 is further configured to convert the requirement keywords into a first embedding vector, and perform a first scene matching from the scene graph library based on the first embedding vector to obtain a first number of basic scenes.

[0130] In one embodiment, the scene recognition module 320 is further configured to convert the extended keywords into a second embedding vector, and perform a second scene matching from the scene graph library based on the second embedding vector to obtain a second number of extended scenes.

[0131] In one embodiment, the scene recognition module 320 is further configured to obtain the text block subsets associated with each scene from the scene graph library, calculate the average similarity based on the first embedding vector and the embedding vector set of the associated text block subsets of each scene, sort each scene from high to low according to the average similarity, and select the first number of scenes before sorting as the base scene.

[0132] In one embodiment, the resource recommendation module 330 is further configured to extract the first keyword of the candidate resource, the second keyword of the mapping scenario, and the third keyword of the existing resource; find the corresponding element values ​​between the first keyword, the second keyword, and the third keyword in the preset keyword weight matrix in the mapping scenario; and construct the compatibility matrix, the matching matrix, and the cost matrix based on the found element values.

[0133] In one embodiment, the demand information acquisition module 310 is further configured to acquire the enterprise's digital asset data, extract keywords from the digital asset data to obtain existing resource keywords, match the existing resource keywords with a preset resource classification dictionary, determine the type and attribute parameters of the existing resources, and complete the identification of existing resources.

[0134] In one embodiment, such as Figure 4 As shown, the above-mentioned device also includes: a scene graph generation module 340, used for extracting keywords; constructing a text block graph structure; constructing a keyword graph; and integrating the keyword graphs. The nodes, edges, and weight matrices form a scenario graph library containing hierarchical relationships of "industry → business activities → scenarios → resources → pain points → value tags".

[0135] In one embodiment, such as Figure 5 The diagram illustrates the structure of an electronic device used to implement an embodiment of this application. The electronic device includes a central processing unit (CPU) 501, which can perform various appropriate actions and processes based on a program stored in a read-only memory (ROM) 502 or a program loaded from a storage portion 508 into a random access memory (RAM) 503. The RAM 503 also stores various programs and data required for the operation of the electronic device. The CPU 501, ROM 502, and RAM 503 are interconnected via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0136] The following components are connected to I / O interface 505: an input section 506 including a keyboard, mouse, etc.; an output section 507 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 508 including a hard disk, etc.; and a communication section 509 including a network interface card such as a LAN card, modem, etc. The communication section 509 performs communication processing via a network such as the Internet. A drive 510 is also connected to I / O interface 505 as needed. A removable medium 511, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 510 as needed so that computer programs read from it can be installed into storage section 508 as needed.

[0137] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer-readable medium carrying instructions that, in such embodiments, can be downloaded and installed from a network via communication section 509, and / or installed from removable medium 511. When the instructions are executed by central processing unit (CPU) 501, the various method steps described in this application are performed.

[0138] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0139] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for recommending enterprise resources based on scene graphs, characterized in that, The method includes: Obtain enterprise resource demand information and identify demand keywords from the resource demand information; Based on the aforementioned requirement keywords, a first number of basic scenarios are matched from the scenario graph library; Based on the weight matrix of keywords preset in the scene graph of the basic scene, the keywords are expanded to form extended keywords corresponding to the required keywords; Based on the extended keywords, a second number of extended scenarios are matched from the scenario graph library, and a mapping scenario matching the resource requirement information is formed according to the extended scenarios and the basic scenarios. Extract a set of candidate resources corresponding to the resource demand information from the mapped scenario; Extract the first keyword of the candidate resource, the second keyword of the mapping scenario, and the third keyword of the existing resource. Find the corresponding element values ​​between the first keyword, the second keyword, and the third keyword in the preset keyword weight matrix in the mapping scenario. Construct a compatibility matrix, a matching matrix, and a cost matrix based on the found element values. Based on the compatibility matrix, matching matrix, and cost matrix, the recommended resources to be pushed to the enterprise are calculated. The compatibility matrix element values ​​in The following calculation method was used to obtain candidate resources. Associated text block subsets and candidate resources For related text block subsets, calculate the first average similarity between two text block subsets. Based on the weight matrix representation and element value and Calculate and The compatibility between them, as the element value The matching matrix element values ​​in The following calculation method was used to obtain candidate resources. Associated text block subsets and mapping scenarios For related text block subsets, calculate the second average similarity between two text block subsets. Based on the weight matrix representation and element value and Calculate and The degree of matching between them is used as the element value. The cost matrix element values ​​in The following calculation method was used to obtain candidate resources. Associated text block subsets and existing resources For related text block subsets, calculate the third average similarity between two text block subsets. Based on the weight matrix representation and element value and Calculate and The value between them, as element value .

2. The method according to claim 1, characterized in that, The step of matching a first number of basic scenarios from the scene graph library based on the demand keywords includes: converting the demand keywords into a first embedding vector, and performing a first scene matching from the scene graph library based on the first embedding vector to obtain a first number of basic scenarios; The step of matching a second number of extended scenes from the scene graph library based on the extended keywords includes: converting the extended keywords into a second embedding vector, and performing a second scene matching from the scene graph library based on the second embedding vector to obtain a second number of extended scenes.

3. The method according to claim 2, characterized in that, The step of performing a first scene matching from the scene graph library based on the first embedding vector to obtain a first number of basic scenes includes: Obtain the text block subsets associated with each scene from the scene graph library, and calculate the average similarity based on the first embedding vector and the embedding vector set of the associated text block subsets of each scene; The scenes are sorted from high to low according to their average similarity, and the scene with the highest similarity among them is selected as the base scene.

4. The method according to claim 1, characterized in that, The element values ​​in the compatibility matrix are positively correlated with the corresponding element values ​​found; the element values ​​in the matching matrix are positively correlated with the corresponding element values ​​found; and the element values ​​in the cost matrix are negatively correlated with the corresponding element values ​​found.

5. The method according to any one of claims 1 to 4, characterized in that, Before extracting the first keyword of candidate resources, the second keyword of the mapping scenario, and the third keyword of existing resources, the following is also included: Acquire the enterprise's digital asset data, extract keywords from the digital asset data, and obtain existing resource keywords; The existing resource keywords are matched with a preset resource classification dictionary to determine the type and attribute parameters of the existing resources, thus completing the identification of existing resources.

6. A scene graph-based enterprise resource recommendation device, characterized in that, The device includes: The demand information acquisition module is used to acquire the enterprise's resource demand information and identify demand keywords from the resource demand information; The scene recognition module is used to match a first number of basic scenes from the scene graph library based on the demand keywords; expand the keywords according to the weight matrix of the keywords of the basic scenes in the scene graph to form extended keywords corresponding to the demand keywords; match a second number of extended scenes from the scene graph library based on the extended keywords; and form a mapping scene that matches the resource demand information according to the extended scenes and the basic scenes. The resource recommendation module is used to extract a set of candidate resources corresponding to the resource demand information from the mapping scenario; extract the first keyword of the candidate resources, the second keyword of the mapping scenario, and the third keyword of the existing resources; find the corresponding element values ​​between the first keyword, the second keyword, and the third keyword in the preset keyword weight matrix in the mapping scenario; construct a compatibility matrix, a matching matrix, and a cost matrix based on the found element values; and calculate the recommended resources to be pushed to the enterprise based on the compatibility matrix, the matching matrix, and the cost matrix. The compatibility matrix element values ​​in The following calculation method was used to obtain candidate resources. Associated text block subsets and candidate resources For related text block subsets, calculate the first average similarity between two text block subsets. Based on the weight matrix representation and element value and Calculate and The compatibility between them, as the element value The matching matrix element values ​​in The following calculation method was used to obtain candidate resources. Associated text block subsets and mapping scenarios For related text block subsets, calculate the second average similarity between two text block subsets. Based on the weight matrix representation and element value and Calculate and The degree of matching between them is used as the element value. The cost matrix element values ​​in The following calculation method was used to obtain candidate resources. Associated text block subsets and existing resources For related text block subsets, calculate the third average similarity between two text block subsets. Based on the weight matrix representation and element value and Calculate and The value between them, as element value .

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores executable instructions that, when executed by a processor, cause the processor to perform the method as described in any one of claims 1 to 5.

8. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to perform the method as described in any one of claims 1 to 5.

Images