An enterprise information generation and retrieval method based on AI and a knowledge graph

Abstract

The present application relates to the field of artificial intelligence, in particular to an enterprise information generation and retrieval method based on AI and knowledge graph, the present application can accurately locate the core elements of user demand by receiving the target retrieval sentence input by the user, extracting each type of entity, avoiding the ambiguity of traditional keyword retrieval, obtaining the attribute similarity score, the relationship similarity comprehensive score and the semantic similarity comprehensive score of each enterprise through analysis, ensuring that the retrieval result is consistent with the basic face of user demand, constructing dynamic weight rules, analyzing the correlation score of each enterprise and the target retrieval sentence of the user, obtaining the sorting result of each enterprise information and returning it to the user, realizing the accurate, flexible and intelligent enterprise retrieval experience.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and more specifically, to a method for generating and retrieving enterprise information based on AI and knowledge graphs. Background Technology

[0002] With the rapid development of big data and artificial intelligence technologies, enterprise data is experiencing explosive growth, encompassing structured operational data, semi-structured log files, and unstructured documents, patents, and more. Traditional information retrieval methods are no longer sufficient to meet enterprises' needs for in-depth data analysis and intelligent decision-making. Against this backdrop, artificial intelligence optimizes retrieval logic through semantic understanding and machine learning, while knowledge graphs integrate multi-source heterogeneous data by leveraging their entity relationship modeling advantages. The fusion of these two technologies provides a new direction for enterprise information retrieval. By constructing an enterprise-level knowledge network, not only can efficient and accurate information matching be achieved, but also implicit relationships between data can be uncovered, helping enterprises to understand market trends, optimize resource allocation, and gain a competitive edge in the fierce business competition.

[0003] Current enterprise information retrieval technologies generally suffer from low levels of intelligence and weak data correlation analysis capabilities. For example, Chinese patent application number 202411420593.7 discloses an enterprise information data retrieval system. This system is used to quickly retrieve data in the data storage module based on user input. The permission granting module can realize information retrieval, keyword extraction, data compression, and integrate data storage, permission control, index optimization, and external data supplementation functions. It provides enterprises with a comprehensive, efficient, and secure information retrieval platform, facilitating rapid information retrieval by internal employees and improving the response speed of information retrieval.

[0004] However, the above-mentioned patent has the following problems: First, the solution is based only on Boolean logic operators to combine search terms, without mentioning fuzzy search or semantic search, and cannot handle natural language semantics (such as synonyms), which may lead to incomplete search coverage, easy to miss key information, and inability to handle semantic ambiguity.

[0005] Second, the external data supplementation in this solution relies solely on TF-IDF keyword matching, without establishing logical relationships between data, making it impossible to identify implicit associations, automatically update keyword weights, and require manual maintenance of the thesaurus, which may lead to association errors. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides a method for generating and retrieving enterprise information based on AI and knowledge graphs, which can effectively solve the problems mentioned in the prior art.

[0007] The objective of this invention can be achieved through the following technical solution: an enterprise information generation and retrieval method based on AI and knowledge graph, the method comprising the following steps: S1. receiving the target retrieval statement input by the user, and extracting entities of various types from it respectively.

[0008] S2. Compare the extracted entities of each type with the entities of each enterprise, and analyze the attribute similarity score, relationship similarity score, and semantic similarity score of each enterprise.

[0009] S3. Construct dynamic weight rules, and analyze the relevance scores of each enterprise to the user's target search statement based on the attribute similarity score, relationship similarity comprehensive score, and semantic similarity comprehensive score of each enterprise. In this way, the ranking results of each enterprise's information are obtained and returned to the user.

[0010] Preferably, the specific method of step S1 is as follows:

[0011] The system receives a target search statement input by a user, the target search statement containing at least one enterprise-related feature or condition.

[0012] The target search statement is segmented and tagged with parts of speech. The segmentation tool is used to split it into a sequence of words, and each word is tagged with a part of speech tag.

[0013] The word segmentation results are traversed, and various types of entities related to enterprises are extracted from the word set. The extracted entities are then stored in the management database. These types of entities include basic entities, relational entities, and semantic entities.

[0014] The basic entities include company name, address, establishment time, and industry; the relational entities include relational paths and relational topology; and the semantic entities include textual descriptive terms and numerical information.

[0015] Preferably, the specific content of step S2 is as follows: S21. Retrieve a preset knowledge graph from the management database, and extract information data of each enterprise from it. The information data is stored in the knowledge graph in the form of nodes and edges. The nodes represent the entities of each type, and the edges represent the relationships between the entities of each type.

[0016] S22. Based on the similarity scores of each enterprise's name, address, establishment time, and industry, the attribute similarity scores of each enterprise are obtained.

[0017] S23. Based on the relationship path similarity score and relationship topology similarity score of each enterprise, the comprehensive relationship similarity score of each enterprise is obtained.

[0018] S24. Based on the textual semantic similarity score and numerical semantic similarity score of each enterprise, the comprehensive semantic similarity score of each enterprise is obtained.

[0019] Preferably, the specific analysis method for the name similarity scores of each enterprise is as follows: if the target search statement contains an enterprise name entity, the enterprise name in the search content is compared with the standard name, abbreviation, and acronym of each enterprise in the knowledge graph.

[0020] Set the company name in the target search query as string A, and the company's standard name / abbreviation / abbreviation in the knowledge graph as string B. Create a two-dimensional table, with the horizontal axis representing the position of each character in string B and the vertical axis representing the position of each character in string A. Each cell in the table represents the minimum number of operations required to convert the characters in string A into the characters in string B.

[0021] Traverse the lengths from 1 to A and from 1 to B, calculate the edit distance between the two strings based on whether the characters are equal, and thus obtain the name similarity score for each company.

[0022] Preferably, the attribute similarity score of each enterprise further includes: if the target search statement contains an address dimension entity, the enterprise address of each enterprise and the address information extracted from the target search statement are decomposed according to the level of province, city, district and street, the matching degree of the same level is compared, and each level is assigned a corresponding weight, so as to calculate the address similarity score of each enterprise.

[0023] If the target search query contains an entity with the establishment time dimension, the establishment time of each enterprise and the time information involved in the target search query are uniformly converted into a timestamp format. The time difference between the two timestamps is calculated, and the time difference is divided into different intervals. Each interval corresponds to a fixed similarity score. Based on the time difference between the two timestamps, the establishment time similarity score of each enterprise is obtained.

[0024] If the target search query contains industry-specific entities, a pre-built industry classification system is used to map the industry to which each company belongs and the industry keywords in the target search query to the corresponding nodes in the industry classification system. Based on the distance between the nodes in the industry classification system, the industry similarity score of each company is calculated.

[0025] The name, address, establishment date, and industry similarity score of each enterprise are weighted and summed according to the pre-set attribute dimension weights to obtain the attribute similarity score of each enterprise.

[0026] Preferably, the specific method for analyzing the industry similarity scores of each enterprise is as follows: industries are divided into levels, with the top level being the major category, and then further subdivided into specific fields to construct a tree-like hierarchical industry classification framework. Each node represents a specific industry category, and weights are assigned to nodes at different levels.

[0027] For industry keywords in the target search query, directly search the industry classification framework of the tree structure to see if there is a completely matching node. If no match can be found, expand using synonyms and industry terms to find the closest bottom-level node.

[0028] The positions of the industry node and keyword node of each enterprise in the classification framework are determined. The minimum hierarchical interval between the industry node of each enterprise and the industry keyword node in the target search statement in the classification framework is obtained. The shortest path length is calculated based on this, and the industry similarity score of each enterprise is obtained.

[0029] Preferably, the specific analysis method for the comprehensive score of the relationship similarity of each enterprise is as follows: if the target retrieval statement contains entity relationship information, analyze the relationship path of enterprise information in the knowledge graph, obtain all relationship paths from the enterprise entity to other related entities, and construct the corresponding relationship path based on the entity relationship information extracted from the target retrieval statement. Compare the two sets of relationship paths, obtain the total length of the same relationship path, and obtain the relationship path similarity score by dividing it by the total length of all obtained relationship paths. The relationship path similarity score of each enterprise is calculated in this way.

[0030] If the target search query contains entity relationship information, the relationship topology of enterprise information in the knowledge graph is abstracted into the connection relationship of nodes and edges of the graph. For the entity relationships involved in the target search query, the corresponding abstract topology is constructed. The graph matching algorithm is used to calculate the similarity between the two topologies and obtain the relationship topology similarity score of each enterprise.

[0031] The relationship path similarity score and relationship topology similarity score of each enterprise are weighted and summed according to the pre-set relationship similarity dimension weights to obtain the comprehensive relationship similarity score of each enterprise.

[0032] Preferably, the specific analysis method for the comprehensive semantic similarity score of each enterprise is as follows: convert the text content in the enterprise information and the text information in the target retrieval statement of each enterprise into vector representations, and obtain the text semantic similarity score of each enterprise by calculating the cosine similarity between the vectors.

[0033] For the numerical attributes involved in the enterprise information and target search statement of each enterprise, the difference ratio between the numerical values ​​involved in the enterprise information and the target search statement is calculated. According to business needs, the difference ratio is divided into multiple intervals, each interval corresponds to a different similarity, and the numerical semantic similarity score of each enterprise is obtained by comparing the difference ratio.

[0034] The textual semantic similarity scores and numerical semantic similarity scores of each enterprise are weighted and summed according to the pre-set semantic similarity dimension weights to obtain the comprehensive semantic similarity score of each enterprise.

[0035] Preferably, the dynamic weighting rule constructs a weighting function based on user search history, search scenario, and enterprise information update frequency, and dynamically adjusts the weights of the comprehensive score of attribute similarity, comprehensive score of relationship similarity, and comprehensive score of semantic similarity of each enterprise in the final matching degree calculation.

[0036] Preferably, step S3 specifically involves: weighting and summing the attribute similarity scores, relation similarity scores, and semantic similarity scores of each enterprise's information according to dynamically adjusted weights to obtain the relevance score between each enterprise and the user's target search statement.

[0037] The information of each enterprise is sorted in descending order based on the relevance score between each enterprise and the user's target search query, and the sorting results are returned to the user.

[0038] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: First, the present invention can quickly match the corresponding attributes in the enterprise database by receiving the target search statement input by the user and extracting various types of entities from it, thereby improving the search response speed.

[0039] Second, by analyzing the attribute similarity score, relationship similarity score, and semantic similarity score of each enterprise, this invention can understand the user's true intention, accurately match the basic information of enterprises, and explore the enterprise relationship network, ensuring that the search results are consistent with the fundamentals of the user's needs.

[0040] Third, this invention constructs dynamic weight rules to analyze the relevance scores of each enterprise to the user's target search statement, obtains the ranking results of each enterprise's information and returns them to the user, ensuring that the results are both comprehensive and tailored to the user's priorities. Attached Figure Description

[0041] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0042] Figure 1This is a schematic diagram of the method flow of the present invention.

[0043] Figure 2 for Figure 1 The flowchart for step S2.

[0044] Figure 3 for Figure 2 A flowchart of the industry similarity scores for each enterprise in step S22. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Please see Figure 1 As shown, the present invention provides an enterprise information generation and retrieval method based on AI and knowledge graph. The method includes the following steps: S1. Receive the target retrieval statement input by the user and extract entities of various types from it.

[0047] The specific method of step S1 is as follows: receiving a target search statement input by the user, wherein the target search statement contains at least one enterprise-related feature or condition.

[0048] The target search statement is segmented and tagged with parts of speech. The segmentation tool is used to split it into a sequence of words, and each word is tagged with a part of speech tag.

[0049] The word segmentation results are traversed, and various types of entities related to enterprises are extracted from the word set. The extracted entities are then stored in the management database. These types of entities include basic entities, relational entities, and semantic entities.

[0050] The basic entities include company name, address, establishment time, and industry; relational entities include relational paths and relational topology; and semantic entities include descriptive text terms and numerical information. By extracting various types of entities from the target search statement, the core elements of user needs can be accurately located, avoiding the ambiguity of traditional keyword searches, facilitating rapid matching of corresponding attributes in the company database, and improving search response speed.

[0051] S2. Compare the extracted entities of each type with the entities of each enterprise, and analyze the attribute similarity score, relationship similarity score, and semantic similarity score of each enterprise.

[0052] Please see Figure 2As shown, the specific content of step S2 is as follows: S21. Retrieve a preset knowledge graph from the management database and extract information data of each enterprise from it. The information data is stored in the knowledge graph in the form of nodes and edges. The nodes represent the entities of each type, and the edges represent the relationships between the entities of each type.

[0053] S22. Based on the name, address, establishment time, and industry similarity score of each enterprise, the attribute similarity score of each enterprise is obtained; the matching degree between the basic attributes of the enterprises and the search statement is compared to ensure that the search results are consistent with the basic needs of users.

[0054] S23. Based on the relationship path similarity score and relationship topology similarity score of each enterprise, the comprehensive relationship similarity score of each enterprise is obtained; the relationship between enterprises (such as upstream and downstream of the supply chain, investment cooperation, competitors, etc.) is analyzed, and the search dimensions are expanded by combining the relational entities in the search statement.

[0055] S24. Based on the textual semantic similarity score and numerical semantic similarity score of each enterprise, the comprehensive semantic similarity score of each enterprise is obtained; natural language processing technology is used to analyze the deep semantic meaning of the search statement and identify synonyms, near-synonyms and domain-specific expressions.

[0056] The specific analysis method for the name similarity scores of each enterprise is as follows: if the target search statement contains an enterprise name entity, the enterprise name in the search content is compared with the standard name, abbreviation, and acronym of each enterprise in the knowledge graph.

[0057] Set the company name in the target search query as string A, and the company's standard name / abbreviation / abbreviation in the knowledge graph as string B. Create a two-dimensional table, with the horizontal axis representing the position of each character in string B and the vertical axis representing the position of each character in string A. Each cell in the table represents the minimum number of operations required to convert the characters in string A into the characters in string B.

[0058] Traverse the lengths from 1 to A and from 1 to B, calculate the edit distance between the two strings based on whether the characters are equal, and thus obtain the name similarity score for each company.

[0059] It should be noted that the specific method for analyzing the edit distance between the two strings is as follows: Let the company name in the target search statement be string A (the name to be matched).

[0060] Let string B (reference name) represent the standard name / abbreviation / acronym of the enterprise in the knowledge graph.

[0061] Construct a two-dimensional array ,in This indicates the preceding text of string A. A substring of characters, This indicates the preceding part of string B. A substring of 100 characters.

[0062] The value indicates that the string A is preceded by... Before converting the characters to string B The minimum number of editing operations required to create a character.

[0063] When one of the strings is empty, the edit distance is the length of the other string: (Prepare string A) (Delete one character to get an empty string) (Insert string B before empty string) (characters).

[0064] Traversal (from 1 to )and (from 1 to ), calculated according to the following rules If the current character is , (No action required, directly inherit the value of the top left corner).

[0065] If the current characters are not equal, take the minimum value among the following three operations: .

[0066] replace: (Put A) The character is replaced with B. (characters).

[0067] delete: (Delete A) (characters).

[0068] insert: (Insert the first line of B into A) (number of characters)

[0069] The value of the bottom right corner of the matrix This is the edit distance between string A and string B. .

[0070] The attribute similarity scores for each enterprise also include: if the target search statement contains an address dimension entity, the enterprise address and the address information extracted from the target search statement are decomposed according to the levels of province, city, district, and street, the matching degree of the same level is compared, and each level is assigned a corresponding weight, thereby calculating the address similarity score of each enterprise; through hierarchical decomposition, misjudgments caused by differences in address descriptions are avoided, and the accuracy of address matching is improved.

[0071] If the target search query contains an entity with an establishment time dimension, the establishment time of each enterprise and the time information involved in the target search query are uniformly converted into a timestamp format. The time difference between the two timestamps is calculated, and the time difference is divided into different intervals. Each interval corresponds to a fixed similarity score. Based on the time difference between the two timestamps, the establishment time similarity score of each enterprise is obtained. The timestamp format eliminates the differences in different time representations, ensuring the uniqueness and accuracy of time calculation.

[0072] If the target search query contains industry-specific entities, a pre-built industry classification system is used to map the industry to which each company belongs and the industry keywords in the target search query to the corresponding nodes in the industry classification system. Based on the distance of the nodes in the industry classification system, the industry similarity score of each company is calculated. The industry classification system ensures the standardization of industry matching and avoids misjudgments caused by the confusion caused by companies filling in their own industry names.

[0073] The name, address, establishment date, and industry similarity score of each enterprise are weighted and summed according to the pre-set attribute dimension weights to obtain the attribute similarity score of each enterprise; to avoid the bias of a single dimension, the core needs of users are highlighted by weight allocation.

[0074] Please see Figure 3 As shown, the specific analysis method for the industry similarity scores of each enterprise is as follows: industries are divided into levels, with the top level being the major category, and then further subdivided into specific fields to construct a tree-like hierarchical industry classification framework. Each node represents a specific industry category, and weights are assigned to nodes at different levels. Through hierarchical classification, fuzzy industry terms are accurately located to the bottom-level nodes, avoiding matching errors caused by confusion in the industry names reported by enterprises. The tree structure can intuitively reflect the hierarchical relationship between industries, making it easy to calculate cross-level correlation.

[0075] For industry keywords in the target search query, the system directly searches the industry classification framework of the tree structure for a complete match. If no match is found, it expands the search using synonyms and industry terms to find the closest bottom-level node. The direct matching mechanism ensures accurate search efficiency, while the synonym expansion solves the problem of the diversity of natural language expressions and avoids the omission of enterprises due to differences in word choice.

[0076] The positions of the industry node and keyword node of each enterprise in the classification framework are determined. The minimum hierarchical interval between the industry node of each enterprise and the industry keyword node in the target search statement in the classification framework is obtained. The shortest path length is calculated to obtain the industry similarity score of each enterprise. The industry relevance is converted into a calculable value by the hierarchical interval number, avoiding the ranking bias caused by subjective judgment.

[0077] It should be noted that the formula for calculating the shortest path length is as follows: ,in These represent the level numbers of nodes A and B, respectively. This represents the nearest common ancestor node, which is the deepest ancestor node shared by two nodes in the tree (i.e., the highest-level node where their paths intersect).

[0078] The specific analysis method for the comprehensive score of relationship similarity among enterprises is as follows: If the target search statement contains entity relationship information, analyze the relationship paths of enterprise information in the knowledge graph, obtain all relationship paths from the enterprise entity to other related entities, and construct corresponding relationship paths based on the entity relationship information extracted from the target search statement. Compare the two sets of relationship paths, obtain the total length of the same relationship paths, and obtain the relationship path similarity score by dividing it by the total length of all obtained relationship paths. Calculate the relationship path similarity score for each enterprise in this way. By limiting the path length and filtering the relationship type, the search scope is quickly narrowed.

[0079] If the target retrieval statement contains entity relationship information, the relationship topology of enterprise information in the knowledge graph is abstracted into the connection relationship of nodes and edges of the graph. For the entity relationships involved in the target retrieval statement, the corresponding abstract topology is constructed. The graph matching algorithm is used to calculate the similarity between the two topologies and obtain the relationship topology similarity score of each enterprise. The topology similarity can capture non-explicit relationship features. If the target retrieval does not explicitly mention "competitive relationship" but implicitly requires enterprises to be "in the same competitive network", the system can identify enterprises in the same competitive cluster through the clustering coefficient (the degree of interconnection between the enterprise and its partners).

[0080] It should be noted that the specific analysis method of the graph matching algorithm is as follows: compare the node attributes in the query pattern graph with the node attributes in the enterprise subgraph to obtain the node attribute similarity, verify the compatibility between the relationships between nodes in the query pattern graph and the corresponding relationships between nodes in the enterprise subgraph to obtain the relationship matching degree, and calculate the topological structure similarity score between the query pattern graph and the enterprise subgraph based on the node attribute similarity and the relationship matching degree.

[0081] The relationship path similarity score and relationship topology similarity score of each enterprise are weighted and summed according to the pre-set relationship similarity dimension weights to obtain the comprehensive relationship similarity score of each enterprise. Relationship path similarity focuses on the accuracy of specific relationship chains, while topology similarity focuses on the macro-match of network form. The combination of the two can avoid the problem of "local matching but overall incompatibility".

[0082] The specific analysis method for the comprehensive semantic similarity score of each enterprise is as follows: convert the text content in the enterprise information and the text information in the target search statement of each enterprise into vector representations, and obtain the text semantic similarity score of each enterprise by calculating the cosine similarity between the vectors.

[0083] It should be noted that the formula for the cosine similarity is: Where A represents the vector representation of the target search statement, and B represents the vector representation of the enterprise text. Represents the vector of the first 1 eigenvalue, .

[0084] For the numerical attributes involved in the enterprise information and target search statement of each enterprise, the difference ratio between the numerical values ​​involved in the enterprise information and the target search statement is calculated. According to business needs, the difference ratio is divided into multiple intervals, each interval corresponds to a different similarity. The numerical semantic similarity score of each enterprise is obtained by comparing the difference ratio. The difference ratio eliminates the influence of differences in numerical units and magnitudes, and uniformly uses relative values ​​to measure the degree of matching.

[0085] The textual semantic similarity scores and numerical semantic similarity scores of each enterprise are weighted and summed according to the pre-set semantic similarity dimension weights to obtain the comprehensive semantic similarity score of each enterprise. The textual semantics cover the qualitative description of the enterprise, and the numerical semantics cover the quantitative indicators. The combination of the two can avoid the one-sidedness of a single data type.

[0086] S3. Construct dynamic weight rules, and analyze the relevance scores of each enterprise to the user's target search statement based on the attribute similarity score, relationship similarity comprehensive score, and semantic similarity comprehensive score of each enterprise. In this way, the ranking results of each enterprise's information are obtained and returned to the user.

[0087] The dynamic weighting rule constructs a weighting function based on user search history, search scenario, and enterprise information update frequency. It dynamically adjusts the weights of the comprehensive score of attribute similarity, comprehensive score of relationship similarity, and comprehensive score of semantic similarity of each enterprise in the final matching degree calculation. According to the user's historical behavior, the weight of address and industry attributes is automatically increased, and enterprises that match the user's regional and industry preferences are displayed first, reducing the user's screening cost.

[0088] It should be noted that the weighting function is... ,in Indicates the first A comprehensive similarity score, , representing the comprehensive score of attribute similarity, the comprehensive score of relation similarity, and the comprehensive score of semantic similarity, respectively. To retrieve historical feature vectors for users, To retrieve scene feature vectors, This is a feature vector representing the frequency of enterprise information updates. This is a weight calculation function based on multiple factors.

[0089] The construction steps of the multi-factor-based weight calculation function are as follows: Obtain the user's search history feature vector, search scenario feature vector, and enterprise information update frequency feature vector respectively; and assign a similarity score to each feature vector. Calculate the contribution of each basic weight, and obtain the temporary weight by merging the contributions of each basic weight.

[0090] It should be noted that the basic weight contribution can be set based on industry experience or obtained through a limited number of experimental data. The specific process involves first collecting user search history data (such as search keywords and click records), search scenario data (such as search intent type and device terminal), and enterprise information update frequency data (such as the number of attribute changes and the frequency of relationship network changes). Then, the correlation coefficients between the above data and attribute similarity, relationship similarity, and semantic similarity are calculated. Next, regression analysis or logistic regression analysis is used to determine the contribution of each data factor to different similarity scores. Finally, through normalization, the contribution is converted into corresponding weights, and the sum of all similarity weights is 1. This dynamically adjusts the weight of each similarity score in the final matching degree calculation.

[0091] The specific content of step S3 is as follows: by weighting and summing the comprehensive scores of attribute similarity, relationship similarity, and semantic similarity of each enterprise's information according to dynamically adjusted weights, the relevance score between each enterprise and the user's target search statement is obtained.

[0092] The information of each enterprise is sorted in descending order based on the relevance score between each enterprise and the user's target search query, and the sorting results are returned to the user. The score sorting makes highly relevant enterprises clear at a glance, and users can directly focus on a specified number of enterprises instead of traversing all results.

[0093] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention, which are still covered within the protection scope of the present invention.

Claims

1. A method for generating and retrieving enterprise information based on AI and knowledge graphs, characterized in that, Includes the following steps: S1. Receive the target search statement input by the user and extract entities of each type from it; S2. Compare the extracted entities of each type with the entities of each enterprise, and analyze the attribute similarity score, relationship similarity score, and semantic similarity score of each enterprise respectively. The attribute similarity scores of each enterprise include calculating the industry similarity score of each enterprise if the target search statement contains an entity related to the industry dimension. The specific method is as follows: industries are divided into levels, with the top level being the major category, and then further subdivided into specific fields to construct a tree-like hierarchical industry classification framework. Each node represents a specific industry category, and weights are assigned to nodes at different levels. For industry keywords in the target search query, directly search for a matching node in the hierarchical industry classification framework. If no match is found, expand using synonyms and industry terms to find the closest bottom-level node. Determine the position of the enterprise's industry node and keyword node in the classification framework. Obtain the minimum hierarchical interval between the enterprise's industry node and the industry keyword node in the target search query in the classification framework. Calculate the shortest path length to obtain the industry similarity score for each enterprise. S3. Construct dynamic weight rules, and analyze the relevance scores of each enterprise to the user's target search statement based on the attribute similarity score, relationship similarity comprehensive score, and semantic similarity comprehensive score of each enterprise. In this way, the ranking results of each enterprise's information are obtained and returned to the user. The dynamic weighting rule is based on user search history, search scenario, and enterprise information update frequency to construct a weighting function, and dynamically adjusts the weights of the comprehensive score of attribute similarity, comprehensive score of relationship similarity, and comprehensive score of semantic similarity of each enterprise in the final matching degree calculation. The weighting function is: ,in Indicates the first A comprehensive similarity score, , representing the comprehensive score of attribute similarity, the comprehensive score of relation similarity, and the comprehensive score of semantic similarity, respectively. To retrieve historical feature vectors for users, To retrieve scene feature vectors, This is a feature vector representing the frequency of enterprise information updates. This is a weight calculation function based on multiple factors; The construction steps of the multi-factor-based weight calculation function are as follows: Obtain the user's search history feature vector, search scenario feature vector, and enterprise information update frequency feature vector respectively; and assign a similarity score to each feature vector. Calculate the contribution of each basic weight, and obtain the temporary weight by merging the contributions of each basic weight; The basic weight contribution is set based on industry experience or obtained through a limited number of experimental data. The specific process is as follows: first, collect user search history data, search scenario data, and enterprise information update frequency data; then, calculate the correlation coefficients between the above data and attribute similarity, relationship similarity, and semantic similarity; next, use regression analysis or logistic regression analysis to determine the contribution of each data factor to different similarity scores; finally, through normalization processing, convert the contribution into corresponding weights, and the sum of the weights of each similarity is 1, thereby dynamically adjusting the weight of each similarity comprehensive score in the final matching degree calculation.

2. The enterprise information generation and retrieval method based on AI and knowledge graphs according to claim 1, characterized in that: The specific method for step S1 is as follows: Receive a target search statement input by a user, wherein the target search statement contains at least one enterprise-related feature or condition; The target retrieval statement is segmented and tagged with parts of speech. The segmentation tool is used to split it into a sequence of words, and each word is tagged with a part of speech tag. The word segmentation results are traversed, and various types of entities related to enterprises are extracted from the vocabulary set. The extracted entities are then stored in the management database. These entities include basic entities, relational entities, and semantic entities. The basic entities include company name, address, establishment time, and industry; the relational entities include relational paths and relational topology; and the semantic entities include textual descriptive terms and numerical information.

3. The enterprise information generation and retrieval method based on AI and knowledge graphs according to claim 1, characterized in that: The specific content of step S2 is as follows: S21. Retrieve a preset knowledge graph from the management database, and extract information data of each enterprise from it. The information data is stored in the knowledge graph in the form of nodes and edges. Nodes represent the various types of entities, and edges represent the relationships between the various types of entities. S22. Attribute similarity scores for each enterprise are obtained based on the enterprise's name, address, establishment date, and industry similarity score. S23. Based on the relationship path similarity score and relationship topology similarity score of each enterprise, the comprehensive relationship similarity score of each enterprise is obtained; S24. Based on the textual semantic similarity score and numerical semantic similarity score of each enterprise, the comprehensive semantic similarity score of each enterprise is obtained.

4. The enterprise information generation and retrieval method based on AI and knowledge graphs according to claim 1, characterized in that: The specific analysis method for the name similarity scores of the aforementioned enterprises is as follows: If the target search query contains the entity of company name, the company name in the search content will be compared with the standard name, abbreviation, and acronym of each company in the knowledge graph; Set the company name in the target search statement as string A, and the company standard name / abbreviation / abbreviation in the knowledge graph as string B. Create a two-dimensional table, with the horizontal axis representing the position of each character in string B and the vertical axis representing the position of each character in string A. Each cell in the table represents the minimum number of operations required to convert the characters of string A into the characters of string B. Traverse the lengths from 1 to A and from 1 to B, calculate the edit distance between the two strings based on whether the characters are equal, and thus obtain the name similarity score for each company.

5. The enterprise information generation and retrieval method based on AI and knowledge graphs according to claim 1, characterized in that: The attribute similarity scores for each enterprise also include: If the target search statement contains an address dimension entity, the enterprise addresses of each enterprise and the address information extracted from the target search statement are broken down according to the level of province, city, district and street. The matching degree of the same level is compared, and each level is assigned a corresponding weight. Then, the address similarity score of each enterprise is calculated. If the target search statement contains an entity with the establishment time dimension, the establishment time of each enterprise and the time information involved in the target search statement are uniformly converted into a timestamp format. The time difference between the two timestamps is calculated, and the time difference is divided into different intervals. Each interval corresponds to a fixed similarity score. Based on the time difference between the two timestamps, the establishment time similarity score of each enterprise is obtained. If the target search statement contains entities related to the industry dimension, the industry to which each enterprise belongs and the industry keywords in the target search statement are mapped to the corresponding nodes in the industry classification system using a pre-built industry classification system. Based on the distance of the nodes in the industry classification system, the industry similarity score of each enterprise is calculated. The name, address, establishment date, and industry similarity score of each enterprise are weighted and summed according to the pre-set attribute dimension weights to obtain the attribute similarity score of each enterprise.

6. The enterprise information generation and retrieval method based on AI and knowledge graphs according to claim 1, characterized in that: The specific analysis method for the comprehensive score of the relationship similarity among the enterprises is as follows: If the target retrieval statement contains entity relationship information, analyze the relationship paths of enterprise information in the knowledge graph, obtain all relationship paths from the enterprise entity to other related entities, and construct corresponding relationship paths based on the entity relationship information extracted from the target retrieval statement. Compare the two sets of relationship paths, obtain the total length of the same relationship paths, and obtain the relationship path similarity score by dividing it by the total length of all obtained relationship paths. In this way, the relationship path similarity score of each enterprise is calculated. If the target search statement contains entity relationship information, the relationship topology of enterprise information in the knowledge graph is abstracted into the connection relationship of nodes and edges of the graph. For the entity relationships involved in the target search statement, the corresponding abstract topology is constructed. The graph matching algorithm is used to calculate the similarity between the two topologies and obtain the relationship topology similarity score of each enterprise. The relationship path similarity score and relationship topology similarity score of each enterprise are weighted and summed according to the pre-set relationship similarity dimension weights to obtain the comprehensive relationship similarity score of each enterprise.

7. The enterprise information generation and retrieval method based on AI and knowledge graphs according to claim 1, characterized in that: The specific analysis method for the comprehensive semantic similarity score of each enterprise is as follows: The text content in the enterprise information and the text information in the target search statement of each enterprise are converted into vector representations. By calculating the cosine similarity between the vectors, the text semantic similarity score of each enterprise is obtained. For the numerical attributes involved in the enterprise information and target search statement of each enterprise, the difference ratio between the numerical values ​​involved in the enterprise information and the target search statement is calculated. According to business needs, the difference ratio is divided into multiple intervals, each interval corresponds to a different similarity, and the numerical semantic similarity score of each enterprise is obtained by comparing the difference ratio. The textual semantic similarity scores and numerical semantic similarity scores of each enterprise are weighted and summed according to the pre-set semantic similarity dimension weights to obtain the comprehensive semantic similarity score of each enterprise.

8. The enterprise information generation and retrieval method based on AI and knowledge graphs according to claim 1, characterized in that: The specific content of step S3 is as follows: By weighting and summing the attribute similarity scores, relational similarity scores, and semantic similarity scores of each enterprise's information according to dynamically adjusted weights, the relevance score between each enterprise and the user's target search statement is obtained. The information of each enterprise is sorted in descending order based on the relevance score between each enterprise and the user's target search query, and the sorting results are returned to the user.

Images