An industry structure map construction method and system

Abstract

A kind of industrial structure atlas construction method, comprising: obtaining multi-source data, the multi-source data includes multiple perspectives;According to the multi-source data, construct atlas;According to the extraction rule of preestablished, information extraction is carried out;The extracted information is aligned with the constructed atlas;Obtain entity alignment industrial structure atlas.The present application expands industrial atlas scale by obtaining multi-source data, multi-source data includes multiple perspectives, to expand the breadth and depth of atlas;By intelligently aligning the extracted information with the constructed atlas, the efficiency of industrial structure atlas construction is improved, without relying on expert experience, reducing time cost and economic cost, reducing construction cost.

Description

Technical Field

[0001] This invention belongs to the field of knowledge graph technology, specifically relating to a method and system for constructing an industrial structure graph. Background Technology

[0002] The products within the technical field of this invention mainly include web search engines, knowledge graph visualization, industry knowledge graphs, industry classification, intelligent question answering, knowledge bases, and intelligent recommendations. Current development of knowledge graphs is primarily focused on intelligence and scaling; however, in terms of intelligence, the quality of entity alignment in knowledge graphs still needs improvement. Regarding scaling, knowledge graphs still face certain technical challenges in ultra-large-scale fusion, requiring targeted design based on different application scenarios.

[0003] Industry structure mapping is a data representation method used to describe the subdivisions of an industry. It can assist in industry analysis and provide data support from an industry structure perspective for establishing a unified national market. Currently, most industry structures are defined and customized by industry experts. Due to limitations imposed by human factors, the breadth and depth of these industry structure maps are also limited. Existing methods for manually constructing maps rely on expert experience, resulting in high time and economic costs, and the scale of the constructed industries is also limited. Existing automated methods for constructing industry structure maps tend to produce relatively simple structures, generally showing hierarchical relationships between industries. Furthermore, the entity alignment process requires retraining vectors, which incurs time, human, and economic costs. Because existing methods compare all vectors, they are also difficult to support the construction of ultra-large-scale industry maps. Summary of the Invention

[0004] (I) Purpose of the Invention

[0005] The purpose of this invention is to provide a method and system for constructing an industry structure map that can expand the scale of the industry map, improve the efficiency of industry structure map construction, and reduce the construction cost.

[0006] (II) Technical Solution

[0007] To address the above problems, the present invention provides a method for constructing an industrial structure map, comprising:

[0008] Acquire multi-source data, which includes multiple perspectives;

[0009] Based on the multi-source data, a map is constructed;

[0010] Information is extracted according to the preset extraction rules;

[0011] Aligning the extracted information with the constructed graph to form entities includes:

[0012] The constructed graph includes multiple nodes, and the multiple nodes are partitioned to obtain multiple partitions;

[0013] The extracted information includes information nodes, and the partition to which the information node belongs is obtained according to a preset prediction algorithm;

[0014] Calculate the similarity between the information node and all nodes in the partition to which it belongs;

[0015] Based on the similarity, the information nodes are aligned with the nodes in the constructed graph according to the preset entity alignment rules.

[0016] Obtain an industry structure map with entity alignment.

[0017] In another aspect of the present invention, preferably, the plurality of perspectives includes at least: an industry perspective, a product perspective, and a technology perspective.

[0018] In another aspect of the present invention, preferably, the preset prediction algorithm for obtaining the partition to which the information node belongs includes:

[0019] Calculate the similarity between the information node and each of the multiple partitions:

[0020]

[0021] Among them, s i v represents the similarity of the current partition. T A vector representing an information node. N represents the set of all nodes in the current partition. x The vectors from the current viewpoint, and the viewpoint set is represented as (1,2,…,x,…,X);

[0022] The region with the highest similarity score is selected as the region to which the information node belongs.

[0023] p = argmax([s1, s2, ..., s...) i , ..., s P ]

[0024] Where p represents the highest similarity score among all partitions, and Argmax represents taking the highest score.

[0025] In another aspect of the present invention, preferably, the preset prediction algorithm for obtaining the partition to which the information node belongs further includes:

[0026] Calculate the set N of all nodes in the current partition. x Vector from the current perspective

[0027]

[0028] N x = [n1, n2, ..., n i ,...]

[0029]

[0030] Where, the current node n i The corresponding vector is v represents any vector, ∈ = 1e-12, and Norm represents normalization.

[0031] In another aspect of the present invention, preferably, the similarity between the computational information node and all nodes in the partition to which it belongs includes:

[0032] Calculate the current node n i vector:

[0033]

[0034]

[0035] N i =[n i1 n i2 ,…,n ik ,…,n iK ]

[0036] in, Represented as the current node n i The vector, n i The set of neighbor nodes is N i , Indicates the current node n i neighbor node n ik The vector s ik Represents n i With n ik The shortest path between [0, 1], where ε is a hyperparameter and takes values ​​between [0, 1]. This represents the comprehensive coefficient.

[0037] In another aspect of the present invention, preferably, the similarity between the computational information node and all nodes in the partition to which it belongs further includes:

[0038] Computational information node w j vector

[0039]

[0040]

[0041] W j =[wj1 w j2 ,…,w jl ,…,w jL ]in, Represented as information node w j The vector, w j The set of neighbor nodes is W j , Information node w j neighbor node w jl The vector.

[0042] In another aspect of the present invention, preferably, the preset entity alignment rule includes:

[0043]

[0044]

[0045] in, To be related to the current node n i Information nodes aligned to entities.

[0046] In another aspect of the present invention, preferably, the preset extraction rules include:

[0047] If the extracted information is structured text, the preset extraction rule is rule extraction.

[0048] If the extracted information is unstructured text, the preset extraction rule is to use deep learning for extraction.

[0049] In another aspect of the present invention, preferably, the construction of the atlas includes:

[0050] Determine the professional field and scope of the map ontology;

[0051] List the important terms in the atlas ontology;

[0052] Define the relationships between nodes in the graph ontology;

[0053] Define the perspective of the node.

[0054] In another aspect, preferably, an industrial structure mapping system includes:

[0055] Acquisition module: Acquires multi-source data, which includes multiple perspectives;

[0056] Construction module: Constructs a map based on the multi-source data;

[0057] Extraction module: Extracts information according to preset extraction rules;

[0058] Entity alignment module: Aligns the extracted information with the constructed graph, including:

[0059] The constructed graph includes multiple nodes, and the multiple nodes are partitioned to obtain multiple partitions;

[0060] The extracted information includes information nodes, and the partition to which the information node belongs is obtained according to a preset prediction algorithm;

[0061] Calculate the similarity between the information node and all nodes in the partition to which it belongs;

[0062] Based on the similarity, the information nodes are aligned with the nodes in the constructed graph according to the preset entity alignment rules.

[0063] Obtain an industry structure map with entity alignment.

[0064] (III) Beneficial Effects

[0065] The above-described technical solution of the present invention has the following beneficial technical effects:

[0066] This invention expands the scale, breadth, and depth of industry graphs by acquiring multi-source data, encompassing multiple perspectives. It employs multi-source data fusion combined with targeted partitioning methods to provide data and technical support for constructing ultra-large-scale industry graphs. Compared to existing automated methods for constructing industry graphs, this approach produces graphs that are larger in scale and more comprehensive in information. Without targeted partitioning, all nodes would require computation, resulting in extremely high overhead and strain on computers. The invention proposes a graph entity alignment method to improve the accuracy of graph representation. Furthermore, it eliminates the need for separate training for entity alignment, reducing labor costs and increasing efficiency. Ultimately, this approach enhances the quality of industry graph construction while reducing costs and increasing efficiency. Attached Figure Description

[0067] Figure 1 This is an overall flowchart of one embodiment of the present invention;

[0068] Figure 2 This is a diagram of one embodiment of the present invention. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0070] The accompanying drawings illustrate a layer structure according to an embodiment of the present invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0071] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0072] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0073] Example 1

[0074] A method for constructing an industrial structure map. Figure 1 An overall flowchart of one embodiment of the present invention is shown, as follows: Figure 1 As shown, it includes:

[0075] Acquiring multi-source data, which includes multiple perspectives; the specific number and content of perspectives are not limited here. Optionally, in this embodiment, the multiple perspectives include at least: an industry perspective, a product perspective, and a technology perspective; the technology perspective mainly constructs a graph structure related to industry technology, and its data sources mainly include, but are not limited to, patents, existing discipline classifications, papers, grant applications, news websites, etc.; the product perspective mainly comes from company websites, e-commerce websites, information released by listed companies, and reports released by relevant consulting institutions; the industry perspective mainly collects domestic and international industry structure classifications, such as the "National Economic Industry Classification," "National Key Supported High-tech Fields," and "Global Industry Classification Standard." Compilation Methods, etc.; For example, marine engineering equipment manufacturing belongs to both the national economic classification and the strategic emerging industries. The two data sources for marine engineering equipment manufacturing are both from an industry perspective. In this embodiment, the multi-source data includes multiple perspectives, which can expand the breadth and depth of the map and increase its scale.

[0076] Based on the multi-source data, a graph is constructed; the specific content of the graph is not limited here, but optionally, in this embodiment, the construction of the graph includes:

[0077] Determine the professional fields and scope of the graph ontology; list the important terms in the graph ontology; define the nodes and their relationships in the graph ontology; define the perspective of the nodes.

[0078] Optionally, in this embodiment, Figure 2 A graph of one embodiment of the present invention is shown, such as Figure 2 As shown, graphs are constructed from industry, product, and technology perspectives. Nodes in the graph corresponding to the industry perspective are filled with points, as shown below. Figure 2 As shown, circles filled with dots represent specific industries. From an industry perspective, the corresponding relationships are partial inclusion and inclusion. Partial inclusion means that some of the industry content of the current child node is consistent with the node's classification, while some content is inconsistent. For example, marine engineering equipment manufacturing is a child node partially included in wind power generation other related equipment and materials manufacturing. However, some, but not all, of the content of marine engineering equipment manufacturing falls under the scope of wind power generation other related equipment and materials manufacturing; some content of marine engineering equipment manufacturing does not belong to wind power generation other related equipment and materials manufacturing. Inclusion means that the entire current industry is covered by the current parent node's industry scope.

[0079] From a product perspective, the nodes in the corresponding graph are those filled with vertical lines, such as... Figure 2 As shown, circles filled with vertical lines represent products or parts thereof. From a product perspective, the corresponding relationships are structural inclusion, category inclusion, and other similar relationships. Structural inclusion primarily refers to the relationship between a complete piece of equipment and the parts required to make up that equipment. For example, in a robot, the structure includes the robotic arm, rotating device, power unit, and control system. Category inclusion includes different categories; for example, in the robot category, welding robots, underwater robots, and firefighting robots are included. Here, inclusion refers to products included from an industry perspective; inclusion from an industry perspective refers to sub-industries included within an industry.

[0080] From a technical perspective, the corresponding technical node is the node whose content is filled with horizontal lines, such as... Figure 2 As shown, circles filled with horizontal lines represent the relationships of technology inclusion and related technology from a technical perspective. Related technology refers to the technologies involved in a product or industry. Technology inclusion refers to the inclusion relationship in technology.

[0081] For example, "Marine Engineering Equipment Manufacturing" is a node from an industry perspective, and its sub-nodes are also nodes from an industry perspective. The relationship between it and its sub-nodes includes both "inclusion" and "partial inclusion." Specifically, "Metal Cutting and Welding Equipment Manufacturing" has some industries under "Marine Engineering Equipment Manufacturing," so its relationship with "Marine Engineering Equipment Manufacturing" is "partial inclusion"; "Oil Floating Engineering Structures" entirely belongs to "Marine Engineering Equipment Manufacturing," so its relationship with "Marine Engineering Equipment Manufacturing" is inclusion.

[0082] "Metal cutting and welding equipment manufacturing" is a node from an industry perspective, while its sub-nodes are nodes from a product perspective. Therefore, the relationship between "metal cutting and welding equipment manufacturing" and its sub-nodes is one of "inclusion". "Metal cutting and welding equipment manufacturing" and its sub-nodes, such as "electric welding machine" and "gas welding frame", are examples of this inclusion relationship.

[0083] "Welding machine" is a node from a product perspective. As can be seen from the given example, its child nodes are also nodes from a product perspective. Therefore, their relationship exhibits both "structural inclusion" and "category inclusion." Compared to "welding machine," its child nodes "transformer," "rectifier," and "frame" are components. Therefore, the relationship between "welding machine" and its child nodes "transformer," "rectifier," and "frame" is "structural inclusion." Compared to "welding machine," its child nodes "plasma arc welding machine," "resistance welding machine," and "electron beam" are different types of "welding machine." Therefore, their relationship is "category inclusion."

[0084] The "rectifier" is a node from a product perspective. As shown in the example in the diagram, its child nodes are nodes from a technology perspective. Therefore, its relationship with its child nodes is "technology-related", representing technologies related to the "rectifier".

[0085] "Rectification technology" is a node from a technological perspective. Figure 1 In the given example, both the parent and child nodes are nodes from a technical perspective; therefore, the relationship is "technological inclusion." Specifically, "conversion technology" includes "rectification technology"; and "rectification technology" includes "active rectification," "synchronous rectification," "electromechanical rectification," and "electrolytic rectification," etc.

[0086] Information is extracted according to preset extraction rules. The specific content of the preset extraction rules is not limited here. Optionally, in this embodiment, the preset extraction rules include: the extracted information includes structured text and unstructured text. If the extracted information is structured text, the preset extraction rule is rule-based extraction; if the extracted information is unstructured text, the preset extraction rule is extraction using deep learning. Structured text refers to pre-formatted text that can be used to extract relationships through rules, such as existing industry classification standards. Unstructured text, or free text, is mainly extracted using relation extraction methods in deep learning. Its main purpose is to extract structured information from unstructured text, forming tuple data containing relationships. The information extraction tools can be general-purpose, such as Stanford OpenIE or ClausIE, or custom-trained, such as UIE or UIE-X.

[0087] From an industry perspective, information extraction involves selecting existing national / international classification systems, such as the National Industrial Classification of Economic Activities, the National Key Supported High-Tech Fields, and the Global Industry Classification Standard. We construct the classification system based on existing nodes in standards such as the "Compilation Methodology" and the "International Standard Industrial Classification of All Economic Activities." The nodes in different industry classifications have inclusion and partial inclusion relationships.

[0088] Since most of the data in this layer is structured or semi-structured, the information extraction method can be rule-based.

[0089] For example, "marine engineering equipment industry" and "marine engineering equipment manufacturing" were extracted from the existing industry classification standard "strategic emerging industries" using a rule-based approach.

[0090] Information extraction from a product perspective primarily relies on three sources: First, products included in existing industry classification systems, such as the "Statistical Product Classification Catalog" and the "Commodity and Service Tax Classification Code." Second, data is constructed from relevant nodes searched through search engines. Using the current node as the search keyword, relevant text information is retrieved from relevant search engines, and then data is constructed from this text. Third, relevant products are extracted from the company's official website. Fourth, data is constructed from patents.

[0091] The product information inherent in existing industry classification systems can be extracted primarily using rule-based extraction. However, obtaining product information from search engines and official websites requires deep learning-based information extraction methods.

[0092] For example, the types of welding machines are mainly obtained from the Baidu corpus. The UIE information extraction method is used, the corpus is pre-annotated, and then fed into the model for training and prediction. The "structural inclusion" of welding machines is also obtained using the UIE corpus, which is extracted from the patent corpus.

[0093] Information extraction from a technological perspective involves data from three sources: first, existing subject classifications, using existing subject classification knowledge to construct a technological system; second, constructing a technological system through related technologies in patents; and third, using Baidu's corpus.

[0094] Existing subject classifications represent pre-structured data, which can be extracted using rule-based methods. Patent data, however, is unstructured text and requires information extraction methods derived from deep learning.

[0095] For example, some of the rectifier-related technologies were extracted from patents using UIE information; others were extracted from Baidu search corpora using UIE information. Patent corpora are specialized corpora, while UIE corpora are general corpora, requiring manual annotation before the trained model can be used. Although Baidu corpora are general corpora, annotation can improve the accuracy of extraction; therefore, they were also obtained through annotation-based special training before model extraction.

[0096] The extracted information is aligned with the constructed graph; the specific content of the alignment is not limited here, but optionally, in this embodiment, the specific content of the entity alignment includes:

[0097] The constructed graph includes multiple nodes, which are then partitioned to obtain multiple partitions. The method of node partitioning is not limited; it can be partitioning all nodes or only some nodes. Optionally, in this embodiment, the partitioning is performed on upper-level nodes. First, different index areas are established for the upper-level nodes. When a new node, i.e., an information node, needs to be integrated into the existing graph, a classification algorithm is performed to predict which area the information node is most likely to fall into. Once the area where the information node will fall is predicted, the algorithm will proceed to the second step: finding the most likely mounting point within the current area.

[0098] The extracted information includes information nodes, and the partition to which the information node belongs is obtained according to a preset prediction algorithm;

[0099] The graph is structured into different tables based on the different top-level nodes. Due to the use of multi-source data, the different top-level nodes may have different names, but their meanings are similar. These top-level nodes are then grouped into the same table, with several different partitions, for example... Figure 2In the example, for instance, the "Marine Engineering Equipment Manufacturing" industry is selected; for the data in the same table, it is divided into 3 different sets according to the perspective of industry, product and technology, forming different vector representations.

[0100] For example, those selected under "Marine Engineering Equipment Manufacturing" are at the top level and belong to the set of nodes from an industry perspective, such as "Marine Engineering Equipment Manufacturing", "Metal Cutting and Welding Equipment Manufacturing", and "Oil Floating Engineering Structures".

[0101] From a product perspective, this refers to a set of nodes that are at the top level, such as "parts for welding equipment" and "electric welding machine".

[0102] From a technical perspective, this refers to a set of nodes that are at the top level, such as "conversion technology", "rectification technology", and "inverter technology".

[0103] The preset prediction algorithm for determining the partition to which an information node belongs includes: calculating the similarity between the information node and each of the multiple partitions.

[0104]

[0105] Among them, s i v represents the similarity of the current partition. T A vector representing an information node. N represents the set of all nodes in the current partition. x The vectors from the current viewpoint, and the viewpoint set is represented as (1,2,…,x,…,X);

[0106] The region with the highest similarity score is selected as the region to which the information node belongs.

[0107] p = argmax([s1, s2, ..., s...) i , ..., s P ]

[0108] Where p represents the highest similarity score among all partitions, and Argmax represents taking the highest score.

[0109] Calculate the set N of all nodes in the current partition. x Vector from the current perspective

[0110]

[0111] N x = [n1, n2, ..., n i ,...]

[0112]

[0113] Where, the current node ni The corresponding vector is v represents any vector, ∈ = 1e-12, and Norm represents normalization;

[0114] Assumption:

[0115]

[0116]

[0117]

[0118] but

[0119]

[0120] Let N be the set of nodes from three different perspectives. x The corresponding vectors are respectively represented as The vector corresponding to the information node is v T Then the score of the information node's vector in the current partition is:

[0121]

[0122] Assumption

[0123]

[0124]

[0125]

[0126] v T =[0.5623, 0.7229, 0.4016]

[0127] but

[0128]

[0129]

[0130]

[0131] s i =max([0.9459, 0.9245, 0.8673]) = 0.9459

[0132] Assume there are 3 partitions: partition 0, partition 1, and partition 2. The score for each partition is:

[0133] s0 = 0.9459

[0134] s1 = 0.8236

[0135] s2 = 0.9821

[0136] The information node falls within the partition:

[0137] p=argmax([0.9459, 0.8236, 0.9821])=2

[0138] Calculate the similarity between the information node and all nodes in the partition to which it belongs; including:

[0139] Let the current node in the graph be n. i Its sibling nodes, ancestor nodes, and descendant nodes together constitute its neighbor nodes. Let n i The neighboring nodes are N i :

[0140] N i =[n i1 n i2 ,…,n ik ,…,n iK ]

[0141] Current node n in the graph i neighbor node n ik The corresponding vector is

[0142] Current node n i The vector representation is:

[0143]

[0144]

[0145] Applying attention models to the vector representation of knowledge graphs may present several challenges. Previous attention methods primarily dealt with sequence problems and were insensitive to position, yet position is crucial. Therefore, positional encoding was introduced. However, graphs cannot be viewed as sequences, making existing positional encoding unsuitable for representing graph nodes. This solution improves upon this by using the distance of the shortest path from neighbors to the current node as a metric. Closer distances result in greater influence, while greater distances result in less influence. The optimized approach... for

[0146]

[0147] in, Represented as the current node n i The vector, n i The set of neighbor nodes is N i , Indicates the current node n i neighbor node nik The vector s ik Represents n i With n ik The shortest path between [0, 1], where ε is a hyperparameter and takes values ​​between [0, 1]. Indicates the comprehensive coefficient;

[0148] Computational information node w j vector

[0149]

[0150]

[0151] W j =[w j1 w j2 ,…,w jl ,…,w jL ]

[0152] in, Represented as information node w j The vector, w j The set of neighbor nodes is W j , express, Information Node W j neighbor node w jl ;

[0153] Assumption:

[0154]

[0155]

[0156]

[0157]

[0158] Let ε = 0.9, s i0 s i1 s i2 =1, 2, 3, then

[0159]

[0160] β j1 ,β j2 ,β j3 =0.9740, -0.4744, 0.9370

[0161]

[0162]

[0163] Based on the similarity, the information nodes are aligned with the nodes in the constructed graph according to the preset entity alignment rules.

[0164] The preset entity alignment rules include:

[0165]

[0166]

[0167] in, To be related to the current node n i Information nodes aligned to entities.

[0168] Using the assumptions obtained in the second step Right now:

[0169]

[0170] Assumption:

[0171]

[0172]

[0173]

[0174] but:

[0175] s i1 s i2 s i3 = -0.0094, 0.8045, -0.5621

[0176] Let topk be the top2

[0177]

[0178] That is, n in the knowledge graph i Aligned entities, taking the text at the 1st and 0th positions of the text information.

[0179] Obtain an industry structure map with entity alignment.

[0180] This embodiment constructs an industry map from three perspectives: industry, product, and technology. It reflects a wider range of industry dimensions and a deeper understanding of industry value, expanding the map's scale. Considering the entire industry, the map extracts valuable common characteristics from each industry and studies the relationships between industry, product, and technology. While maintaining a rich industry structure, it can reflect the current state of the entire industry from multiple dimensions, providing multi-dimensional value analysis. This embodiment employs multi-source data fusion combined with targeted partitioning methods to provide data and technical support for the construction of ultra-large-scale industry maps. Compared to existing automated methods for constructing industry maps, this method produces a larger-scale map with broader information. Without targeted partitioning, all nodes would need to be calculated, resulting in extremely high computational costs that are difficult for computers to handle. The targeted partitioning and indexing method can perform vector calculations in the most likely regions, significantly reducing the computational scale. Therefore, it provides technical support for the construction of ultra-large-scale maps. A targeted entity alignment method is proposed to improve the accuracy of graph representation. Furthermore, entity alignment does not require separate training, reducing labor costs and improving efficiency. This approach enhances the quality of industry map construction while reducing costs and increasing efficiency.

[0181] Example 2

[0182] An industry structure mapping system, comprising:

[0183] Acquisition module: Acquires multi-source data, which includes multiple perspectives; the specific number of perspectives and the specific content of the perspectives are not limited here. Optionally, in this embodiment, the multiple perspectives include at least: industry perspective, product perspective and technology perspective.

[0184] Construction module: Constructs a map based on the multi-source data;

[0185] Extraction Module: Extracts information according to preset extraction rules. The specific content of the preset extraction rules is not limited here. Optionally, in this embodiment, the preset extraction rules include: the extracted information includes structured text and unstructured text. If the extracted information is structured text, the preset extraction rule is rule-based extraction; if the extracted information is unstructured text, the preset extraction rule is extraction using deep learning. Structured text refers to pre-formatted text that can be extracted using rules, such as existing industry classification standards. Unstructured text, or free text, is mainly extracted using relation extraction methods in deep learning. Its main purpose is to extract structured information from unstructured text, forming tuple data containing relations. The information extraction tools can be general-purpose, such as Stanford OpenIE or ClausIE, or custom-trained, such as UIE or UIE-X.

[0186] From an industry perspective, information extraction involves selecting existing national / international classification systems, such as the National Industrial Classification of Economic Activities, the National Key Supported High-Tech Fields, and the Global Industry Classification Standard. We construct the classification system based on existing nodes in standards such as the "Compilation Methodology" and the "International Standard Industrial Classification of All Economic Activities." The nodes in different industry classifications have inclusion and partial inclusion relationships.

[0187] Since most of the data in this layer is structured or semi-structured, the information extraction method can be rule-based.

[0188] For example, "marine engineering equipment industry" and "marine engineering equipment manufacturing" were extracted from the existing industry classification standard "strategic emerging industries" using a rule-based approach.

[0189] Information extraction from a product perspective primarily relies on three sources: First, products included in existing industry classification systems, such as the "Statistical Product Classification Catalog" and the "Commodity and Service Tax Classification Code." Second, data is constructed from relevant nodes searched through search engines. Using the current node as the search keyword, relevant text information is retrieved from relevant search engines, and then data is constructed from this text. Third, relevant products are extracted from the company's official website. Fourth, data is constructed from patents.

[0190] The product information inherent in existing industry classification systems can be extracted primarily using rule-based extraction. However, obtaining product information from search engines and official websites requires deep learning-based information extraction methods.

[0191] For example, the types of welding machines are mainly obtained from the Baidu corpus. The UIE information extraction method is used, the corpus is pre-annotated, and then fed into the model for training and prediction. The "structural inclusion" of welding machines is also obtained using the UIE corpus, which is extracted from the patent corpus.

[0192] Information extraction from a technological perspective involves data from three sources: first, existing subject classifications, using existing subject classification knowledge to construct a technological system; second, constructing a technological system through related technologies in patents; and third, using Baidu's corpus.

[0193] Existing subject classifications represent pre-structured data, which can be extracted using rule-based methods. Patent data, however, is unstructured text and requires information extraction methods derived from deep learning.

[0194] For example, some of the rectifier-related technologies were extracted from patents using UIE information; others were extracted from Baidu search corpora using UIE information. Patent corpora are specialized corpora, while UIE corpora are general corpora, requiring manual annotation before the trained model can be used. Although Baidu corpora are general corpora, annotation can improve the accuracy of extraction; therefore, they were also obtained through annotation-based special training before model extraction.

[0195] Entity alignment module: Aligns the extracted information with the constructed graph; the entity alignment of the extracted information with the constructed graph includes:

[0196] The constructed graph includes multiple nodes, and the multiple nodes are partitioned to obtain multiple partitions;

[0197] The extracted information includes information nodes, and the partition to which the information node belongs is obtained according to a preset prediction algorithm; the preset prediction algorithm for obtaining the partition to which the information node belongs includes:

[0198] Calculate the similarity between the information node and each of the multiple partitions:

[0199]

[0200] Among them, s i v represents the similarity of the current partition. T A vector representing an information node. N represents the set of all nodes in the current partition. x The vectors from the current viewpoint, and the viewpoint set is represented as (1,2,…,x,…,X);

[0201] The region with the highest similarity score is selected as the region to which the information node belongs.

[0202] p = argmax([s1, s2, ..., s...) i , ..., s P ]

[0203] Where p represents the highest similarity score among all partitions, and Argmax represents taking the highest score.

[0204] The preset prediction algorithm for obtaining the partition to which the information node belongs also includes:

[0205] Calculate the set N of all nodes in the current partition. x Vector from the current perspective

[0206]

[0207] N x = [n1, n2, ..., n i ,...]

[0208]

[0209] Where, the current node n i The corresponding vector is v represents any vector, ∈ = 1e-12, and Norm represents normalization.

[0210] Calculate the similarity between the information node and all nodes in the partition to which it belongs; including:

[0211] Calculate the current node n i vector:

[0212]

[0213]

[0214] N i =[n i1 n i2 ,…,n ik ,…,n iK ]

[0215] in, Represented as the current node n i The vector, n i The set of neighbor nodes is N i , Indicates the current node n i neighbor node n ik The vector s ik Represents n i With n ikThe shortest path between [0, 1], where ε is a hyperparameter and takes values ​​between [0, 1]. This represents the comprehensive coefficient.

[0216] In another aspect of the present invention, preferably, the similarity between the computational information node and all nodes in the partition to which it belongs further includes:

[0217] Computational information node w j vector

[0218]

[0219]

[0220] W j =[w j1 w j2 ,…,w jl ,…,w jL ]

[0221] in, Represented as information node w j The vector, w j The set of neighbor nodes is W j , express, Information node w j neighbor node w jl ;

[0222] Based on the similarity, the information nodes are aligned with the nodes in the constructed graph according to a preset entity alignment rule; the preset entity alignment rule includes:

[0223]

[0224]

[0225] in, To be related to the current node n i Information nodes aligned to entities.

[0226] Obtain an industry structure map with entity alignment.

[0227] It should be understood that the specific embodiments described above are merely illustrative or explanatory of the principles of the invention and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of the invention should be included within the protection scope of the invention. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

[0228] The present invention has been described above with reference to embodiments thereof. However, these embodiments are merely illustrative and not intended to limit the scope of the invention. The scope of the invention is defined by the appended claims and their equivalents. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of the invention, and all such substitutions and modifications should fall within the scope of the invention.

Claims

1. An industrial structure map construction method, characterized by, The method comprises the following steps: obtaining multi-source data, wherein the multi-source data comprises multiple perspectives, including an industry perspective, a product perspective and a technology perspective; constructing a graph according to the multi-source data, comprising: determining the professional field and category of the graph ontology; listing important terms in the graph ontology; defining nodes and the relationship between nodes in the graph ontology; and defining the perspective of the nodes; performing information extraction according to a preset extraction rule; performing entity alignment on the extracted information and the constructed graph, comprising: the constructed graph comprises multiple nodes, the multiple nodes are partitioned to obtain multiple partitions, and the multiple node partitions are performed on upper-layer nodes to establish different index areas for the upper-layer nodes; the extracted information comprises information nodes, and a preset prediction algorithm is used to obtain the partitions to which the information nodes belong; calculating the similarity between the information nodes and all nodes in the partitions to which the information nodes belong; performing entity alignment on the information nodes and the nodes in the constructed graph according to the similarity and a preset entity alignment rule; obtaining an entity-aligned industrial structure graph; the preset prediction algorithm used to obtain the partitions to which the information nodes belong comprises: the similarity between the information nodes and the multiple partitions is calculated respectively; wherein s i represents the similarity of the current partition, represents the vector of the information node, represents the vector of all nodes N x in the current partition under the current view, the view set is represented as (1,2, , x, , X); the corresponding area with the highest similarity score is selected as the area to which the information nodes belong; wherein p represents the highest similarity score of each partition, and argmax represents the highest score.

2. The industry structure map construction method according to claim 1, characterized by, the preset prediction algorithm used to obtain the partitions to which the information nodes belong further comprises: Compute the vector of the current perspective of all nodes set N within the current partition x :​ where the current node n i The corresponding vector is , denotes an arbitrary vector, , Norm denotes normalization.

3. The industry structure map construction method according to claim 2, characterized by, the calculation of the similarity between the information nodes and all nodes in the partitions to which the information nodes belong comprises: Computes the vector of the current node : wherein, is a vector representing the current node , is a set of neighbor nodes of the current node , is a vector representing a neighbor node nik of the current node , sik is a shortest path between ni and nik, is a hyperparameter taking values between , is a synthesis coefficient.

4. The industry structure map construction method according to claim 3, characterized by, the calculation of the similarity between the information nodes and all nodes in the partitions to which the information nodes belong further comprises: Computing information node of vectors : wherein, a vector of information nodes , a set of neighbor nodes of information node , a neighbor node of information node a vector of information nodes.

5. The industry structure map construction method according to claim 4, characterized by, the preset entity alignment rule comprises: wherein, is an information node aligned with the current node entity.

6. The industry structure map construction method according to claim 1, characterized by, the preset extraction rule comprises: if the extracted information is structured text, the preset extraction rule is rule extraction; if the extracted information is unstructured text, the preset extraction rule is extraction using deep learning.

7. An industry structure map construction system characterized by comprising: The method comprises the following steps: an obtaining module: obtaining multi-source data, wherein the multi-source data comprises multiple perspectives, including an industry perspective, a product perspective and a technology perspective; a constructing module: constructing a graph according to the multi-source data, comprising: determining the professional field and category of the graph ontology; listing important terms in the graph ontology; defining nodes and the relationship between nodes in the graph ontology; and defining the perspective of the nodes; an extracting module: performing information extraction according to a preset extraction rule; an entity alignment module: performing entity alignment on the extracted information and the constructed graph, comprising: the constructed graph comprises multiple nodes, the multiple nodes are partitioned to obtain multiple partitions, and the multiple node partitions are performed on upper-layer nodes to establish different index areas for the upper-layer nodes; the extracted information comprises information nodes, and a preset prediction algorithm is used to obtain the partitions to which the information nodes belong; calculating the similarity between the information nodes and all nodes in the partitions to which the information nodes belong; performing entity alignment on the information nodes and the nodes in the constructed graph according to the similarity and a preset entity alignment rule; obtaining an entity-aligned industrial structure graph; the preset prediction algorithm used to obtain the partitions to which the information nodes belong comprises: the similarity between the information nodes and the multiple partitions is calculated respectively; the corresponding area with the highest similarity score is selected as the area to which the information nodes belong. where s i represents the similarity of the current partition, represents the vector of the information node, represents the vector of all nodes N x in the current partition under the current view, and the view set is represented as (1, 2, , x, , X); Select the corresponding area with the highest similarity score as the area to which the information node belongs: Wherein, p represents the highest similarity score of each partition, and argmax represents the highest score.

Images