Enterprise user portraying method based on knowledge graph

Abstract

The invention discloses an enterprise user portraying method based on a knowledge graph, and relates to the technical field of knowledge graphs. Comprising an automatic generation stage and a manual correction stage, the automatic generation stage comprises the following steps that a user inputs an enterprise name, the enterprise name serves as a core node to obtain an unstructured data source, the unstructured data source comprises a text information base, an image information base, an audio information base and a video information base, a semantic object is extracted from text information, and the semantic object is corrected; the method comprises the following steps: acquiring an image recognition object from image information, analyzing data, and converting a semantic object and the recognition object into nodes and a connection relationship, calculating an associated numerical value and utilizing the same label nodes, carrying out disambiguation on the constructed knowledge graph for multiple times, removing the nodes which do not meet the requirements of a user, and improving the user experience. The user can conveniently and efficiently obtain the enterprise user portrait, the user is allowed to add emphasized labels and emphasized connection relations, and the knowledge graph can be synchronously updated.

Description

Technical Field

[0001] The present invention relates to the field of knowledge graph technology, and specifically to an enterprise user profiling method based on knowledge graph. Background Art

[0002] Knowledge graph is a graph-based data structure that represents knowledge as a collection of nodes and edges. Edges refer to the connection between nodes and are used to describe the relationship between entities. Building enterprise user portraits based on knowledge graphs can quickly transform enterprise-related unstructured data sources into easy-to-view graphical data structures, and use the advantages of computer network data crawling to speed up users' efficiency in obtaining enterprise-related data.

[0003] Existing knowledge graphs only rely on big data acquisition to summarize enterprise user portraits. The advantages of machine crawling data will bring disadvantages. There are only a small amount of valuable values ​​in the large amount of data crawled from the Internet. On the one hand, the general computer methods of acquiring and filtering data cannot filter out valuable data well, resulting in the generated knowledge graph being too large and complicated or not meeting user data requirements. On the other hand, the construction of the knowledge graph cannot obtain user feedback, and the acquisition and filtering methods cannot be automatically changed according to the data requirements focused by the user. Therefore, how to optimize the construction of the knowledge graph of enterprise user portraits is a technical problem that technicians in this field need to solve. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides an enterprise user profiling method based on knowledge graph, which solves the problems raised in the above background technology.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for enterprise user profiling based on knowledge graph, including an automatic generation stage and a manual correction stage, wherein the automatic generation stage includes the following steps: S01: Obtaining unstructured data sources. The user inputs the company name and uses the company name as the core node to obtain unstructured data sources. The unstructured data sources include text information library, image information library, audio information library and video information library. S02: Obtain objects, extract semantic objects from text information, obtain image recognition objects from image information, analyze data, convert semantic objects and recognition objects into nodes and connection relationships, the node content includes time, location, object and event, verify the meaning of the connection relationship between nodes, and establish the first knowledge graph; S03: Data fusion: Analyze the label of each node, calculate the similarity between nodes, fuse the data of the same nodes, and obtain the second knowledge graph after fusion; S04: Entity disambiguation, calculating the correlation values ​​between similar nodes and their adjacent nodes, filtering out nodes that match the enterprise user profile from the second knowledge graph, and obtaining the third knowledge graph after filtering; S05: Secondary disambiguation, counting the path length d5 ​​between nodes with the same label and similar nodes, filtering out nodes that match the enterprise user profile from the third knowledge graph, and obtaining the fourth knowledge graph after filtering; S06: perform third disambiguation, calculate the correlation index f between nodes with the same label and similar nodes, select nodes that match the enterprise user profile from the fourth knowledge graph, obtain the fifth knowledge graph, and stop the automatic generation phase; During the manual correction stage, users automatically generate a focus graph and a correlation graph by adding focus labels and focus connection relationships, and then connect the focus graph and the correlation graph to the fifth knowledge graph to obtain the final knowledge graph.

[0006] Furthermore, obtaining the unstructured data source specifically includes the following steps: S101: Based on the enterprise name input by the user, the text search engine is used to search for the business information, news headlines, enterprise qualifications and project names related to the enterprise name, and all the page contents corresponding to the searched business information, news headlines, enterprise qualifications and project names are copied and included in the text information database. The text search engine is the Apache Lucene Java full-text search framework; S102: Obtain the names of people associated with the enterprise from the industrial and commercial public information, extract the semantic objects in the industrial and commercial information, news headlines, enterprise qualifications and project names in S11, input each semantic object and person's name into an image search engine to search for online images, and download all online images into the image information library. The image search engine is a vector search engine Milvus; S103: input the company name and the person's name into an audio search engine to search for audio information on the Internet, download the audio information and add it into an audio information library, the audio search engine is a SoLoud audio engine; S104: inputting the company name and the person's name into a video search engine to search for video information on the Internet, downloading the video information and adding it into a video information database, wherein the video search engine is a Bing search engine; S105: splitting the video information in the video information library into a plurality of frames of image information and audio information, sequentially adding the split image information into the image information library in the order of the video time axis, adding the audio information split from the video information into the audio information library, and converting all the audio information in the audio information library into text information using a sound-to-text program and adding the text information into the text information library, wherein the sound-to-text program is a Whisper speech recognition model; S106: Delete the video information library and the audio information library, and only keep the text information library and the image information library.

[0007] Furthermore, the recognition object includes a text object and a face object, and obtaining the recognition object from the image information library specifically includes the following steps: S201: Obtain text objects, use a text extraction program to scan all image information to obtain text objects, and add the text objects into a text information library. The text extraction program is a Paddle OCR recognition model; S202: Acquire a face object, scan all image information based on a face recognition algorithm to obtain the face object, the face recognition algorithm is an OpenCV algorithm, and temporarily store the face object for subsequent processing; The specific process of implementing face recognition with OpenCV algorithm is as follows: The acquired image is converted into a grayscale image with a grayscale value range of 0-255. The grayscale image is decomposed into a matrix, and each element in the matrix represents the grayscale value of each pixel in the grayscale image. OpenCV uses the SIFT algorithm for image feature matching. The image classifier is obtained by continuously using the matrix converted from the grayscale image for feature training. The image classifier is used to perform feature similarity calculation to obtain the face object area from the image. The above algorithms are all existing open source technologies that can extract face objects from images and will not be described in detail here. Extracting semantic objects from text information specifically includes the following steps: S203: preprocessing, using a semantic analysis program to filter out semantic objects from the text information library, the semantic analysis program is a LPIR-Parser semantic analysis model; S204: object classification, analyzing the semantic attributes of all semantic objects, and classifying all semantic objects according to the semantic attributes, the categories mainly include time, location, object, event and connection; S205: extract the person's name from the text information, analyze the semantic objects of the contact type, combine with the face object obtained in S202, use the face similarity algorithm to match the person's name with the face object one by one, the face similarity algorithm is the face_recognition face recognition model, search the person's name and simultaneously obtain the public portrait photos related to the person's name, compare the face information obtained in S202 with the portrait photo to calculate the similarity, and bind the face information with a similarity higher than the threshold of 95% to the person's name.

[0008] Furthermore, analyzing the data specifically includes the following steps: S206: converting semantic objects of the categories of time, location, object and event into nodes, converting semantic objects of the category of connection into connection relationships, and connecting all nodes using the connection relationships; S207: marking the face object as a portrait tag, and inputting the portrait tag into the node converted from the person's name. The portrait tag can help the user distinguish different person nodes; S208: Verify the connection relationship. Perform a logical analysis on the connection relationship. If the connection relationship between the two nodes has opposite meanings or a subordinate relationship contradiction, the opposite meaning means that there is both a D relationship and an E relationship between the two nodes, and D and E are antonyms from a semantic analysis. The subordinate relationship contradiction means that the F node and the G node both belong to the H node, and the F node belongs to the G node. There is a subordinate relationship contradiction between the F node and the H node. The content and connection relationship of the two nodes are re-entered into the search engine for search, and the number of results of the connection relationship with opposite meanings or a subordinate relationship contradiction is counted. The connection relationship with a lower number of results is deleted from the two nodes. Relationship verification can improve the data authenticity and accuracy of the knowledge graph and avoid the interference of erroneous data in the search, and then proceed to the next step. S209: All nodes and connection relationships are combined into a first knowledge graph.

[0009] Furthermore, data fusion specifically includes the following steps: S301: assigning a label to each node according to the connection relationship; S302: Mark nodes with the same content as similar nodes, calculate the connection relationship similarity and label similarity of similar nodes, and if condition 1 is satisfied that the connection relationship similarity is greater than a threshold of 95% and condition 2 is satisfied that the label similarity is greater than a threshold of 85%, mark the similar nodes as the same nodes; otherwise, if either condition 1 or condition 2 is not satisfied, mark the two similar nodes in numerical order, and mark different numerical suffixes after the same node content, such as "node KL1" and "node KL2", and so on, to distinguish nodes with the same content; S303: Merge the same nodes, and simultaneously merge the labels and connection relationships to obtain a second knowledge graph.

[0010] Furthermore, entity disambiguation specifically includes the following steps: S401: Mark nodes with the same content in the second knowledge graph as similar nodes. The same content here does not include the numeric suffix of the node content. For example, "node KL1" and "node KL2" are still similar nodes; S402: Counting the association between two similar nodes and the core node, where the association includes the number of connections and the length of connections. The number of connections is the total number of connection paths from the similar nodes to the core node, and the length of connections is the total length of the node paths included in all the number of connections. S403: Multiply the number of connections by the connection length to obtain the correlation value between the similar node and the core node, compare the correlation values ​​of the two similar nodes, remove the similar node with the smaller correlation value from the second knowledge graph, and if the correlation values ​​of the two similar nodes are the same, retain the two similar nodes to obtain the third knowledge graph.

[0011] Furthermore, the secondary disambiguation specifically includes the following steps: S501: Select any two nodes with the same label from the third knowledge graph and mark them as a first reference node and a second reference node respectively; S502: Obtain similar nodes retained in S04, count the shortest length d1 from similar nodes to the first reference node, count the shortest length d2 from similar nodes to the second reference node, count the shortest length d3 from similar nodes passing through the core node to the first reference node, and count the shortest length d4 from similar nodes passing through the core node to the second reference node; S503: Calculate according to the formula The path lengths d5 of the first reference node, the second reference node and the similar node, w1 and w2 are proportional coefficients and satisfy the condition w1+w2=1, w1 and w2 determine the specific values ​​according to the number of nodes in the knowledge graph, compare the path lengths d5 of two similar nodes, and remove the similar nodes with smaller path lengths d5 from the third knowledge graph. If the path lengths d5 of two similar nodes are the same, retain the two similar nodes again to obtain the fourth knowledge graph.

[0012] Furthermore, the third disambiguation specifically includes the following steps: S601: Select any two nodes with the same label from the fourth knowledge graph and mark them as a third reference node and a fourth reference node respectively; S602: Acquire similar nodes retained in S05, count the number of connections n1 from similar nodes to the third reference node, count the number of connections n2 from similar nodes to the fourth reference node, count the number of connections n3 from similar nodes passing through the core node to the third reference node, and count the number n4 from similar nodes passing through the core node to the fourth reference node; S603: Count the total length s1 of the number of connections n1, count the total length s2 of the number of connections n2, count the total length s3 of the number of connections n3, and count the total length s4 of the number of connections n4, where the total length s1 is equal to the sum of the shortest lengths of each of the number of connections n1, and the counting method of s2, s3, and s4 is similar; S604: According to the formula Calculate the association index f of the third reference node, the fourth reference node and the similar node, w3, w4, w5 and w6 are proportional coefficients and satisfy the condition w3+w4+w5+w6=1, w3, w4, w5 and w6 are determined according to the number of nodes in the knowledge graph, compare the association index f of two similar nodes, remove the similar nodes with smaller association index f from the fourth knowledge graph, if the association index f of two similar nodes is the same, retain the two similar nodes again to obtain the fifth knowledge graph.

[0013] Furthermore, the manual correction stage includes the following steps: S07, adding a focus label, the user drags the label of any node to another node, marks the node receiving the new label as a focus node, and deletes the nodes and connection relationships related to the original focus node from the fifth knowledge graph; S08: Reconstruct the nodes, re-search the unstructured data sources related to the focus nodes and labels, and execute S02 to S06 again to obtain a focus graph with the focus nodes as the core; S09: Adding a focus connection relationship, the user copies the connection relationship between any two nodes to another two nodes, marks the copied connection relationship as a focus relationship, and deletes all connection relationships between the two nodes connected by the focus relationship from the fifth knowledge graph; S10: Reconstruct the connection relationship, re-search the unstructured data source related to the focus relationship and the two connected nodes, and execute S02 to S06 again to obtain the association graph with the two connected nodes as the core; S11: Output the knowledge graph, merge the focus graph and the association graph in the manual correction stage with the fifth graph in the automatic generation stage, and connect the focus graph and the association graph to the fifth knowledge graph to obtain the final knowledge graph.

[0014] Furthermore, each time the user adds an emphasis label, the emphasis graph is updated synchronously, and the added emphasis label is taken as the same label. When the emphasis graph is recalculated next time, any two nodes with both emphasis labels are selected in S06 for calculation, so that the subsequent three disambiguations can be closer to the emphasis label selected by the user. Each time the user adds an emphasis connection relationship, the associated graph is updated synchronously.

[0015] The present invention has the following beneficial effects: 1. By calculating the associated values ​​and using nodes with the same label, the constructed knowledge graph can be disambiguated multiple times, and nodes that do not meet user needs can be removed. The huge and complex knowledge graph can be streamlined and purified, making it easier for users to efficiently obtain enterprise user portraits.

[0016] 2. By allowing users to add emphasis labels and emphasis connection relationships, the knowledge graph can be updated synchronously to better adapt to the data needs of different users. At the same time, the emphasis labels added by users can be used in the next automatic generation stage, and the calculation rules of the third-order disambiguation can be adjusted to make the automatic generation stage closer to the data needs emphasized by users, reducing the number of updates to the knowledge graph in the manual correction stage.

[0017] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for describing the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0019] Figure 1 It is a flow chart of the automatic generation phase of the enterprise user portrait method based on the knowledge graph of the present invention; Figure 2 It is a flow chart of the manual correction stage of the present invention. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] See also Figure 1-2 ,The present invention provides a technical solution: a method for enterprise user profiling based on knowledge graph, including an automatic generation stage and a manual correction stage, such as Figure 1 As shown, the automatic generation phase includes the following steps: S01: Obtaining unstructured data sources. The user inputs the company name and uses the company name as the core node to obtain unstructured data sources. The unstructured data sources include text information library, image information library, audio information library and video information library. S02: Obtain objects, extract semantic objects from text information, obtain image recognition objects from image information, analyze data, convert semantic objects and recognition objects into nodes and connection relationships, the node content includes time, location, object and event, verify the meaning of the connection relationship between nodes, and establish the first knowledge graph; S03: Data fusion: Analyze the label of each node, calculate the similarity between nodes, fuse the data of the same nodes, and obtain the second knowledge graph after fusion; S04: Entity disambiguation, calculating the correlation values ​​between similar nodes and their adjacent nodes, filtering out nodes that match the enterprise user profile from the second knowledge graph, and obtaining the third knowledge graph after filtering; S05: Secondary disambiguation, counting the path length d5 ​​between nodes with the same label and similar nodes, filtering out nodes that match the enterprise user profile from the third knowledge graph, and obtaining the fourth knowledge graph after filtering; S06: perform third disambiguation, calculate the correlation index f between nodes with the same label and similar nodes, select nodes that match the enterprise user profile from the fourth knowledge graph, obtain the fifth knowledge graph, and stop the automatic generation phase; During the manual correction stage, users automatically generate a focus graph and a correlation graph by adding focus labels and focus connection relationships, and then connect the focus graph and the correlation graph to the fifth knowledge graph to obtain the final knowledge graph.

[0022] Wherein, obtaining the unstructured data source specifically includes the following steps: S101: Based on the enterprise name input by the user, the text search engine is used to search for the business information, news headlines, enterprise qualifications and project names related to the enterprise name, and all the page contents corresponding to the searched business information, news headlines, enterprise qualifications and project names are copied and included in the text information database. The text search engine is the Apache Lucene Java full-text search framework; S102: Obtain the names of people associated with the enterprise from the industrial and commercial public information, extract the semantic objects in the industrial and commercial information, news headlines, enterprise qualifications and project names in S11, input each semantic object and person's name into an image search engine to search for online images, and download all online images into the image information library. The image search engine is a vector search engine Milvus; S103: input the company name and the person's name into an audio search engine to search for audio information on the Internet, download the audio information and add it into an audio information library, the audio search engine is a SoLoud audio engine; S104: inputting the company name and the person's name into a video search engine to search for video information on the Internet, downloading the video information and adding it into a video information database, wherein the video search engine is a Bing search engine; S105: splitting the video information in the video information library into a plurality of frames of image information and audio information, sequentially adding the split image information into the image information library in the order of the video time axis, adding the audio information split from the video information into the audio information library, and converting all the audio information in the audio information library into text information using a sound-to-text program and adding the text information into the text information library, wherein the sound-to-text program is a Whisper speech recognition model; S106: Delete the video information library and the audio information library, and only keep the text information library and the image information library.

[0023] The recognition objects include text objects and face objects. The process of obtaining the recognition objects from the image information database specifically includes the following steps: S201: Obtain text objects, use a text extraction program to scan all image information to obtain text objects, and add the text objects into a text information library. The text extraction program is a Paddle OCR recognition model; S202: Acquire a face object, scan all image information based on a face recognition algorithm to obtain the face object, the face recognition algorithm is an OpenCV algorithm, and temporarily store the face object for subsequent processing; Extracting semantic objects from text information specifically includes the following steps: S203: preprocessing, using a semantic analysis program to filter out semantic objects from the text information library, the semantic analysis program is a LPIR-Parser semantic analysis model; S204: Object classification, analyzing the semantic attributes of all semantic objects, and classifying all semantic objects according to the semantic attributes. The categories mainly include time, location, object, event and connection. For example, the text information "User A books a flight ticket to City B for User C one day before October 1st" can filter out the data points "User A", "User C", "September 30th", "City B", and "Booking a flight ticket", which can be divided into five categories: "User A" and "User C" are objects, "September 30th" is time, "City A" is location, "Booking a flight ticket" is event, and "User A" and "User C" are connections; S205: extract the person's name from the text information, analyze the semantic objects of the contact type, combine with the face object obtained in S202, use the face similarity algorithm to match the person's name with the face object one by one, the face similarity algorithm is the face_recognition face recognition model, search the person's name and simultaneously obtain the public portrait photos related to the person's name, compare the face information obtained in S202 with the portrait photo to calculate the similarity, and bind the face information with a similarity higher than the threshold of 95% to the person's name.

[0024] The data analysis specifically includes the following steps: S206: converting semantic objects of the categories of time, location, object and event into nodes, converting semantic objects of the category of connection into connection relationships, and connecting all nodes using the connection relationships; S207: marking the face object as a portrait tag, and inputting the portrait tag into the node converted from the person's name. The portrait tag can help the user distinguish different person nodes; S208: Verify the connection relationship, perform logical analysis on the connection relationship, if the connection relationship between the two nodes has opposite meanings or a subordinate relationship contradiction, the opposite meaning means that there is both a D relationship and an E relationship between the two nodes, and D and E are antonyms from a semantic analysis, and the subordinate relationship contradiction means that the F node and the G node both belong to the H node, and the F node belongs to the G node, and there is a subordinate relationship contradiction between the F node and the H node, the content and connection relationship of the two nodes are re-entered into the search engine for search, and the number of results of the connection relationship with opposite meanings or a subordinate relationship contradiction is counted, and the connection relationship with a lower number of results is deleted from the two nodes, and the next step is entered; S209: All nodes and connection relationships are combined into a first knowledge graph.

[0025] The data fusion specifically includes the following steps: S301: assigning a label to each node according to the connection relationship. For example, if the connection relationship between the node "user I" and the node "user J" is a cooperative relationship, the label of the node "user I" is "partner of user J", and the label of the node "user J" is "partner of user I"; S302: Mark nodes with the same content as similar nodes, calculate the connection relationship similarity and label similarity of similar nodes, and if condition 1 is satisfied that the connection relationship similarity is greater than a threshold of 95% and condition 2 is satisfied that the label similarity is greater than a threshold of 85%, mark the similar nodes as the same nodes; otherwise, if either condition 1 or condition 2 is not satisfied, mark the two similar nodes in numerical order, and mark different numerical suffixes after the same node content, such as "node KL1" and "node KL2", and so on, to distinguish nodes with the same content; S303: Merge the same nodes, and simultaneously merge the labels and connection relationships to obtain a second knowledge graph.

[0026] Among them, entity disambiguation specifically includes the following steps: S401: Mark nodes with the same content in the second knowledge graph as similar nodes. The same content here does not include the numeric suffix of the node content. For example, "node KL1" and "node KL2" are still similar nodes; S402: Count the association between two similar nodes and the core node. The association includes the number of connections and the length of the connection. The number of connections is the total number of connection paths from the similar node to the core node. The connection length is the total length of the node paths included in all the number of connections. For example, node M is connected to node N and node O respectively. Node N and node O are both connected to the core node. Node N is connected to node O. The number of connections of node M is 4, which are MN-core node, path length is 2, MO-core node, path length is 2, MNO-core node, path length is 3, MON-core node, path length is 3, and the connection length is 2+2+3+3=10. S403: Multiply the number of connections by the connection length to obtain the correlation value between the similar node and the core node, compare the correlation values ​​of the two similar nodes, remove the similar node with the smaller correlation value from the second knowledge graph, and if the correlation values ​​of the two similar nodes are the same, retain the two similar nodes to obtain the third knowledge graph.

[0027] The secondary disambiguation specifically includes the following steps: S501: Select any two nodes with the same label from the third knowledge graph and mark them as a first reference node and a second reference node respectively; S502: Obtain similar nodes retained in S04, count the shortest length d1 from similar nodes to the first reference node, count the shortest length d2 from similar nodes to the second reference node, count the shortest length d3 from similar nodes passing through the core node to the first reference node, and count the shortest length d4 from similar nodes passing through the core node to the second reference node; S503: Calculate according to the formula The path lengths d5 of the first reference node, the second reference node and the similar node, w1 and w2 are proportional coefficients and satisfy the condition w1+w2=1, w1 and w2 determine the specific values ​​according to the number of nodes in the knowledge graph, compare the path lengths d5 of two similar nodes, and remove the similar nodes with smaller path lengths d5 from the third knowledge graph. If the path lengths d5 of two similar nodes are the same, retain the two similar nodes again to obtain the fourth knowledge graph.

[0028] The three-step disambiguation process specifically includes the following steps: S601: Select any two nodes with the same label from the fourth knowledge graph and mark them as a third reference node and a fourth reference node respectively; S602: Acquire similar nodes retained in S05, count the number of connections n1 from similar nodes to the third reference node, count the number of connections n2 from similar nodes to the fourth reference node, count the number of connections n3 from similar nodes passing through the core node to the third reference node, and count the number n4 from similar nodes passing through the core node to the fourth reference node; S603: Count the total length s1 of the number of connections n1, count the total length s2 of the number of connections n2, count the total length s3 of the number of connections n3, and count the total length s4 of the number of connections n4, where the total length s1 is equal to the sum of the shortest lengths of each of the number of connections n1, and the counting method of s2, s3, and s4 is similar; S604: According to the formula Calculate the association index f of the third reference node, the fourth reference node and the similar node, w3, w4, w5 and w6 are proportional coefficients and satisfy the condition w3+w4+w5+w6=1, w3, w4, w5 and w6 are determined according to the number of nodes in the knowledge graph, compare the association index f of two similar nodes, remove the similar nodes with smaller association index f from the fourth knowledge graph, if the association index f of two similar nodes is the same, retain the two similar nodes again to obtain the fifth knowledge graph.

[0029] Among them, Figure 2 As shown, the manual correction stage includes the following steps: S07, adding a focus label, the user drags the label of any node to another node, marks the node receiving the new label as a focus node, and deletes the nodes and connection relationships related to the original focus node from the fifth knowledge graph; S08: Reconstruct the nodes, re-search the unstructured data sources related to the focus nodes and labels, and execute S02 to S06 again to obtain a focus graph with the focus nodes as the core; S09: Adding a focus connection relationship, the user copies the connection relationship between any two nodes to another two nodes, marks the copied connection relationship as a focus relationship, and deletes all connection relationships between the two nodes connected by the focus relationship from the fifth knowledge graph; S10: Reconstruct the connection relationship, re-search the unstructured data source related to the focus relationship and the two connected nodes, and execute S02 to S06 again to obtain the association graph with the two connected nodes as the core; S11: Output the knowledge graph, merge the focus graph and the association graph in the manual correction stage with the fifth graph in the automatic generation stage, and connect the focus graph and the association graph to the fifth knowledge graph to obtain the final knowledge graph.

[0030] Each time the user adds an emphasis label, the emphasis graph is recalculated and the added emphasis label is used as the same label. When the emphasis graph is recalculated next time, any two nodes with both emphasis labels are selected in S06 for calculation, so that the subsequent three disambiguations can be closer to the emphasis label selected by the user.

[0031] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. The enterprise user profiling method based on knowledge graph is characterized by: It includes the automatic generation stage and the manual correction stage. The automatic generation stage includes the following steps: S01: Obtaining unstructured data sources. The user inputs the company name and uses the company name as the core node to obtain unstructured data sources. The unstructured data sources include text information library, image information library, audio information library and video information library. S02: Obtain objects, extract semantic objects from text information, obtain image recognition objects from image information, analyze data, convert semantic objects and recognition objects into nodes and connection relationships, the node content includes time, location, object and event, verify the meaning of the connection relationship between nodes, and establish the first knowledge graph; S03: Data fusion: Analyze the label of each node, calculate the similarity between nodes, fuse the data of the same nodes, and obtain the second knowledge graph after fusion; S04: Entity disambiguation, calculating the correlation values ​​between similar nodes and their adjacent nodes, filtering out nodes that match the enterprise user profile from the second knowledge graph, and obtaining the third knowledge graph after filtering; S05: Secondary disambiguation, counting the path length d5 ​​between nodes with the same label and similar nodes, filtering out nodes that match the enterprise user profile from the third knowledge graph, and obtaining the fourth knowledge graph after filtering; S06: perform third disambiguation, calculate the correlation index f between nodes with the same label and similar nodes, select nodes that match the enterprise user profile from the fourth knowledge graph, obtain the fifth knowledge graph, and stop the automatic generation phase; During the manual correction stage, users automatically generate a focus graph and a correlation graph by adding focus labels and focus connection relationships, and then connect the focus graph and the correlation graph to the fifth knowledge graph to obtain the final knowledge graph.

2. The enterprise user profiling method based on knowledge graph according to claim 1 is characterized in that: Obtaining unstructured data sources specifically includes the following steps: S101: Based on the enterprise name input by the user, the text search engine is used to search for the business information, news headlines, enterprise qualifications and project names related to the enterprise name, and all the page contents corresponding to the searched business information, news headlines, enterprise qualifications and project names are copied and included in the text information database; S102: Obtain the names of people associated with the enterprise from the industrial and commercial public information, extract the semantic objects in the industrial and commercial information, news headlines, enterprise qualifications and project names in S11, input each semantic object and person's name into an image search engine to search for online images, and download all online images into the image information library; S103: Input the company name and the person's name into an audio search engine to search for audio information on the Internet, and download the audio information to be included in an audio information database; S104: inputting the company name and the person's name into a video search engine to search for video information on the Internet, and downloading the video information into a video information database; S105: splitting the video information in the video information library into a plurality of frames of image information and audio information, sequentially adding the split image information into the image information library in the order of the video time axis, adding the audio information split from the video information into the audio information library, and converting all the audio information in the audio information library into text information using a sound-to-text program and adding the text information into the text information library, wherein the sound-to-text program is a Whisper speech recognition model; S106: Delete the video information library and the audio information library, and only keep the text information library and the image information library.

3. The enterprise user profiling method based on knowledge graph according to claim 1 is characterized in that: The recognition objects include text objects and face objects. The steps of obtaining the recognition objects from the image information library specifically include the following steps: S201: Acquire text objects, use a text extraction program to scan all image information to acquire text objects, and add the text objects into a text information library; S202: Acquire a face object, scan all image information based on a face recognition algorithm to obtain the face object, and temporarily store the face object for subsequent processing; Extracting semantic objects from text information specifically includes the following steps: S203: preprocessing, using a semantic analysis program to filter out semantic objects from a text information database; S204: object classification, analyzing the semantic attributes of all semantic objects, and classifying all semantic objects according to the semantic attributes, the categories mainly include time, location, object, event and connection; S205: extracting the person's name from the text information, analyzing the semantic objects of the contact type, combining the face objects obtained in S202, and using the face similarity algorithm to match the person's name with the face objects one by one.

4. The enterprise user profiling method based on knowledge graph according to claim 3 is characterized in that: Analyzing the data specifically includes the following steps: S206: converting semantic objects of the categories of time, location, object and event into nodes, converting semantic objects of the category of connection into connection relationships, and connecting all nodes using the connection relationships; S207: Marking the face object as a portrait tag, and inputting the portrait tag into the node converted from the person's name; S208: Verify the connection relationship, perform logical analysis on the connection relationship, if the connection relationship between two nodes has opposite meanings or a subordinate relationship contradiction, re-enter the content and connection relationship of the two nodes into the search engine, count the number of results of the connection relationship of the meaning, delete the connection relationship with a lower number of results from the two nodes, and proceed to the next step; S209: All nodes and connection relationships are combined into a first knowledge graph.

5. The enterprise user profiling method based on knowledge graph according to claim 1 is characterized in that: Data fusion specifically includes the following steps: S301: assigning a label to each node according to the connection relationship; S302: Mark nodes with the same content as similar nodes, calculate the connection relationship similarity and label similarity of similar nodes, and if condition 1 is met that the connection relationship similarity is greater than a threshold of 95% and condition 2 is met that the label similarity is greater than a threshold of 85%, mark the similar nodes as the same nodes; otherwise, if either condition 1 or condition 2 is not met, mark the two similar nodes in numerical order, and mark different numerical suffixes after the same node content; S303: Merge the same nodes, and simultaneously merge the labels and connection relationships to obtain a second knowledge graph.

6. The enterprise user profiling method based on knowledge graph according to claim 1 is characterized in that: Entity disambiguation specifically includes the following steps: S401: Mark nodes with the same content in the second knowledge graph as similar nodes; S402: Counting the association between two similar nodes and the core node, where the association includes the number of connections and the length of connections. The number of connections is the total number of connection paths from the similar nodes to the core node, and the length of connections is the total length of the node paths included in all the number of connections. S403: Multiply the number of connections by the connection length to obtain the correlation value between the similar node and the core node, compare the correlation values ​​of the two similar nodes, remove the similar node with the smaller correlation value from the second knowledge graph, and if the correlation values ​​of the two similar nodes are the same, retain the two similar nodes to obtain the third knowledge graph.

7. The enterprise user profiling method based on knowledge graph according to claim 1 is characterized in that: Secondary disambiguation specifically includes the following steps: S501: Select any two nodes with the same label from the third knowledge graph and mark them as a first reference node and a second reference node respectively; S502: Obtain similar nodes retained in S04, count the shortest length d1 from similar nodes to the first reference node, count the shortest length d2 from similar nodes to the second reference node, count the shortest length d3 from similar nodes passing through the core node to the first reference node, and count the shortest length d4 from similar nodes passing through the core node to the second reference node; S503: Calculate according to the formula The path lengths d5 of the first reference node, the second reference node and the similar node, w1 and w2 are proportional coefficients and satisfy the condition w1+w2=1. The path lengths d5 of the two similar nodes are compared, and the similar nodes with smaller path lengths d5 are removed from the third knowledge graph. If the path lengths d5 of the two similar nodes are the same, the two similar nodes are retained again to obtain the fourth knowledge graph.

8. The enterprise user profiling method based on knowledge graph according to claim 1 is characterized in that: The three-step disambiguation process specifically includes the following steps: S601: Select any two nodes with the same label from the fourth knowledge graph and mark them as a third reference node and a fourth reference node respectively; S602: Acquire similar nodes retained in S05, count the number of connections n1 from similar nodes to the third reference node, count the number of connections n2 from similar nodes to the fourth reference node, count the number of connections n3 from similar nodes passing through the core node to the third reference node, and count the number n4 from similar nodes passing through the core node to the fourth reference node; S603: Count the total length s1 of the number of connections n1, count the total length s2 of the number of connections n2, count the total length s3 of the number of connections n3, and count the total length s4 of the number of connections n4, where the total length s1 is equal to the sum of the shortest lengths of each of the number of connections n1, and the counting method of s2, s3, and s4 is similar; S604: According to the formula Calculate the association index f of the third reference node, the fourth reference node and the similar node, w3, w4, w5 and w6 are proportional coefficients and satisfy the condition w3+w4+w5+w6=1, compare the association index f of two similar nodes, remove the similar nodes with smaller association index f from the fourth knowledge graph, if the association index f of two similar nodes is the same, retain the two similar nodes again to obtain the fifth knowledge graph.

9. The enterprise user profiling method based on knowledge graph according to claim 1 is characterized in that: The manual correction phase includes the following steps: S07, adding a focus label, the user drags the label of any node to another node, marks the node receiving the new label as a focus node, and deletes the nodes and connection relationships related to the original focus node from the fifth knowledge graph; S08: Reconstruct the nodes, re-search the unstructured data sources related to the focus nodes and labels, and execute S02 to S06 again to obtain a focus graph with the focus nodes as the core; S09: Adding a focus connection relationship, the user copies the connection relationship between any two nodes to another two nodes, marks the copied connection relationship as a focus relationship, and deletes all connection relationships between the two nodes connected by the focus relationship from the fifth knowledge graph; S10: Reconstruct the connection relationship, re-search the unstructured data source related to the focus relationship and the two connected nodes, and execute S02 to S06 again to obtain the association graph with the two connected nodes as the core; S11: Output the knowledge graph, merge the manual correction stage with the automatic generation stage, and connect the focus graph and the associated graph to the fifth knowledge graph to obtain the final knowledge graph.

10. The enterprise user profiling method based on knowledge graph according to claim 9 is characterized in that: Each time the user adds an emphasis label, the emphasis map is updated synchronously, and the added emphasis label is used as the same label. When the emphasis map is recalculated next time, any two nodes with both emphasis labels are selected in S06 for calculation.

Images