A multi-domain knowledge fusion method based on semantic tree

Abstract

The application discloses a multi-field knowledge fusion method based on semantic trees, and mainly relates to the technical field of knowledge graph in multi-source and multi-field knowledge cross fusion. The application is based on the establishment of semantic trees of multiple fields of knowledge, proposes a corresponding association rule mining method, performs association analysis between the semantic trees, finds the association rules of different fields, and performs mining and calculation of associated elements based on graph operation theory, so that organic fusion of multi-field knowledge can be realized. The application can solve the problems of not easy alignment, association and merging of semantic trees of different fields due to heterogeneity, and can fuse multi-source and multi-field knowledge into an organic whole, so that more comprehensive, unified and cooperative knowledge sharing can be provided, and the application is especially suitable for theoretical and systematic design of a unified technical architecture in a 'knowledge-intensive' field.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of knowledge graph in multi-source and multi-field knowledge cross-fusion, reflecting the correlation and interaction of knowledge systems and elements in various fields, and is particularly suitable for the theoretical and systematic design of unified technical architecture in "knowledge-intensive" fields. BACKGROUND

[0002] In the era of rapid development of advanced scientific technologies such as artificial intelligence, big data, and deep learning, knowledge graph, as an effective tool for knowledge system construction, fusion, and application, can establish corresponding knowledge semantic trees through the integration and integrated processing of large-scale data in various fields and applications, so as to facilitate the extraction of knowledge. To realize the deep fusion of cross-field knowledge, it is necessary to merge and process knowledge semantic trees of different sources and different fields. However, due to the heterogeneity of knowledge semantic trees in different fields, the correlation between them is complex and inconsistent in hierarchical division, which makes it impossible to directly and simply merge them. SUMMARY

[0003] The present application aims to address the shortcomings in the background art and provides a multi-field knowledge fusion method based on semantic trees. Based on the establishment of multiple field semantic trees, the method aligns, correlates, and merges multiple semantic trees through knowledge correlation, making them an organic whole to provide more comprehensive, unified, and collaborative knowledge sharing.

[0004] The technical solution adopted by the present application is as follows: based on the establishment of multiple field knowledge semantic trees, an appropriate correlation rule mining method is proposed, and the correlation between semantic trees is analyzed to find the correlation rules in different fields ， and based on graph operation theory to mine and calculate the correlation elements, realizing the organic fusion of multi-field knowledge. The method includes the following steps:

[0005] Step 1: Based on big data text materials, model and attribute fill for multiple knowledge fields to be studied, and expand the connection between entities to construct a hierarchical knowledge semantic tree reflecting the superior-inferior and subordinate structure of knowledge at different levels;

[0006] Step 2: Based on the establishment of multiple field semantic trees, an appropriate correlation rule mining method is proposed for the knowledge characteristics of different fields, and cross-level and cross-field correlation rules are formed using existing text materials, and an index of correlation strength is given based on this;

[0007] Step 3: Use the FP-growth algorithm to perform correlation analysis on the multi-field knowledge semantic trees to quickly scan and obtain the correlation rules between them;

[0008] Step 4, based on the graph operation theory, the mining and calculation of associated elements are carried out, and the most authoritative node and the nearest secondary authoritative node to the target node are searched from the mass nodes;

[0009] Step 5, the associated rules mined from the knowledge semantic tree are integrated with the knowledge trees in multiple fields, and the characteristics of various fields are fused, and according to the association relationship between elements, a comprehensive knowledge graph system of multi-field fusion is constructed, and the unity and interaction of the main trunk and branches in the knowledge graph reflect the association of the knowledge system and elements in various fields;

[0010] Step 6, a graph database is used for storage.

[0011] Further, in the step 2, the association rules are used to reflect the interdependence and association between one thing and other things, and are used to mine the correlation between valuable data items from a large amount of data;

[0012] The association strength of the association rules is described by support, confidence and lift;

[0013] The extraction process of the association rules is to find all the association rules with support >= minimum support and confidence >= minimum confidence;

[0014] The association rules meeting the conditions are mined in two steps: first, generate the frequent item set, that is, find all the item sets meeting the minimum support, and then generate the rules meeting the minimum confidence based on the frequent item set;

[0015] The correlation between different field knowledge is obtained by association rule mining, and the overall correlation degree of the multi-field knowledge system is directly reflected, thereby laying a theoretical foundation for the establishment of the comprehensive knowledge graph.

[0016] Further, in the step 4, the most authoritative node is determined by using the mutual voting method, if the data amount of the knowledge graph is very large, a block type is used to realize, the PageRank value of each block graph is calculated first, the new graph PageRank value is obtained according to the correlation between the data blocks, and then the authoritative node is analyzed by repeated iteration;

[0017] The multi-feature method based on node attributes and the relationship between nodes is used to analyze the authoritative node by comprehensively analyzing the node attributes and the relationship.

[0018] The above technical scheme has the following advantages:

[0019] 1. The knowledge graph of multi-field fusion is constructed, the knowledge in multiple sources and multiple fields is fused into an organic whole, the theoretical guidance of the unified technical system architecture design is provided, and the comprehensive performance is improved.

[0020] 2、The design method for cross-domain knowledge fusion of the application adopts knowledge graph technology, is based on big data text materials, adopts artificial intelligence data mining mode, enables the machine to understand the meaning of the text, avoids the influence of human subjective factors and the limitation of personal knowledge, and the analysis result is more objective, comprehensive, reasonable, and has higher theoretical feasibility.

[0021] 3、The algorithm selected by the application has relatively short time consumption, high efficiency, low calculation complexity and small workload, and through optimization processing, the technical feasibility is high. ACCURACY

[0022] Figure 1 is a flowchart of the application;

[0023] Figure 2 is a Dijkstra algorithm flowchart. DETAILED DESCRIPTION

[0024] The application will be further described below in combination with the drawings and examples.

[0025] As shown in Figure 1 , the application comprises the following steps:

[0026] (1) Construction of multiple field knowledge semantic trees

[0027] Based on big data text materials, modeling and attribute filling are carried out for multiple knowledge fields to be studied, and the connection between entities is expanded, a hierarchical knowledge semantic tree is constructed, which can clearly reflect the superior-inferior and subordinate structure of different levels of knowledge, so that the field knowledge can be sorted out from multiple angles.

[0028] Different domain original text knowledge is converted into a structured semantic tree for representation, so that the text principle can be decomposed, just like the trunk and its branches, seeking its leaves and details downward, so as to form hierarchical and interlayer relationships. In this way, the original knowledge is converted into a deeper and more abstract semantic tree, which is conducive to learning transfer.

[0029] (2) Mining of association rules

[0030] On the basis of establishing multiple field semantic trees, according to the knowledge characteristics of different fields, a suitable association rule mining method is proposed, the existing text materials are used to form cross-level and cross-domain association rules, and on this basis, an index of association degree strength is given.

[0031] The association rules reflect the interdependence and correlation between one thing and other things, and are used to mine the correlation between valuable data items from a large amount of data. The present application is based on the correlation between different vocabularies, and mines the correlation between different layers of the same tree and different trees, finds the correlation between different vocabularies between layers or between trees, and thus realizes the correlation mining between different fields and different layers.

[0032] The correlation strength of the association rules can be described by support and confidence. There are three key concepts in the analysis of the association rules: support, confidence and lift.

[0033] The support refers to the proportion of the occurrence times of several associated data in the total data set:

[0034]

[0035] Wherein, Y is the former of the association rules, or is called the precondition; X is the latter of the association rules, or is called the correlation result; when the item set Y occurs, the item set X also has a certain probability of occurrence. The confidence refers to the probability of the occurrence of one data after the occurrence of another data, or the conditional probability of the data:

[0036]

[0037] The lift indicates the probability of containing X under the condition of containing Y, and the ratio of the probability of the occurrence of X to the total probability of the occurrence of X:

[0038]

[0039] Based on the above definitions, the extraction process of the association rules is to find all the association rules with support >= minimum support and confidence >= minimum confidence. In order to quickly mine the association rules meeting the conditions, two steps can be performed: firstly, generating the frequent item set, i.e. finding all the item sets meeting the minimum support, and then generating the rules, generating the rules meeting the minimum confidence on the basis of the frequent item set, and the generated rules are called strong rules.

[0040] The correlation between different field knowledge can be obtained by the association rule mining, and the overall correlation degree of the multi-field knowledge system is directly reflected, which lays a theoretical foundation for the establishment of the comprehensive knowledge graph.

[0041] (3) Operation of the correlation analysis

[0042] After comprehensive comparison of various algorithms, the FP-growth algorithm is used for association analysis of the knowledge semantic tree in multiple fields, and the association rules between them can be quickly scanned. Specifically, the divide-and-conquer strategy is adopted, and after the first scan, the frequent item set in the database is compressed into a frequent pattern tree, while the association information is still retained, then the tree is divided into some conditional libraries, each library is related to a frequent item set with a length of 1, and then the conditional libraries are mined. When the amount of original data is very large, the partitioning method can be combined to make the tree can be put into the main memory.

[0043] In this way, the semantic frequent pattern mining of the previously established knowledge semantic tree is completed, and in this process, the correlation degree index between modules-support is generated, and by setting the support threshold, the association rules in different fields can be found.

[0044] (4) Mining and calculation of associated elements

[0045] Based on the graph operation theory, the most authoritative node (important node) and the next authoritative node closest to the target node (shortest path) are found from the massive nodes.

[0046] The determination of the most authoritative node needs to pass through the most authoritative node analysis step, and the authority of the node is analyzed from the knowledge graph, and the mutual voting method is adopted. The idea of this method is derived from the PageRank idea, and the PageRank refers to the webpage pointed by more high-quality webpages, which has a higher high-quality probability. If the data amount of the knowledge graph is very large, in order to reduce the algorithm overhead, the block type can be used to realize the algorithm, the PageRank value of each block graph is calculated, the correlation between the data blocks is obtained, the new graph PageRank value is obtained, and the authoritative node is analyzed by repeatedly iterating. The multi-feature method based on node attributes and node relationship can be used to analyze the authoritative node by comprehensively analyzing the node attributes and relationship.

[0047] The shortest path algorithm adopted by the present application is the Dijkstra algorithm, which is described with reference to Figure 2 The core idea is to expand layer by layer from the starting point until the end point is reached. The algorithm steps are as follows:

[0048] a) At the beginning, S only contains the original point v, and the distance is 0, and U represents the vertex set opposite to S.

[0049] b) Select a vertex k with the minimum distance from U, and add k to the S set.

[0050] c) Take k as another original point, modify the shortest distance to the original point for each vertex in U, and if the distance to k is less than the distance to v, modify the original distance to a smaller value.

[0051] d) Repeat steps 2, 3 until all vertices are added to the set S.

[0052] Wherein, S is the node set of finding the shortest path, U is the node set of not finding the shortest path, the set S and U constitute the whole node.

[0053] (5) Establishment of comprehensive knowledge graph

[0054] For the original data from different channels or data feature attributes processed in different ways, the corresponding knowledge tree can be constructed based on the respective knowledge system. By integrating the knowledge trees of multiple fields and the association rules mined for the knowledge semantic tree, the characteristics of various fields can be integrated, and according to the association relationship between elements, a comprehensive knowledge graph system of multi-field integration can be constructed, and the unity and interaction of the main trunk and branches in the knowledge graph reflect the association of the knowledge systems and elements of various fields.

[0055] (6) Storage of knowledge graph

[0056] For the comprehensive knowledge graph constructed, a suitable method needs to be selected for storage. At present, the storage of knowledge graph can be divided into table structure-based storage and graph structure-based storage. In order to make the multi-field integrated knowledge graph have better visualization, directly and accurately reflect its internal system structure, and be beneficial to the deep mining and reasoning of knowledge, it is suggested to use a graph database for storage.

Claims

1. A multi-domain knowledge fusion method based on semantic tree, characterized in that, The specific steps are as follows: Step 1, based on big data, model and attribute fill in multiple knowledge fields that need to be studied, and expand the relationship between entities to build a hierarchical knowledge semantic tree, reflecting the superior-inferior and subordinate structure of different levels of knowledge; Step 2, based on the establishment of multiple field semantic trees, propose a correlation rule mining method suitable for the knowledge characteristics of different fields, use existing text materials to form cross-level and cross-field association rules, and give the index of association strength on this basis; Step 3, use the FP-growth algorithm to perform association analysis on the multi-field knowledge semantic tree, and quickly scan to obtain the association rules between them; Step 4, based on graph operation theory, mine and calculate the associated elements, find the most authoritative node and the nearest secondary authoritative node from the target node in the massive nodes; Step 5, integrate the knowledge trees of multiple fields and the association rules mined from the knowledge semantic trees, fuse the characteristics of various fields, construct a comprehensive knowledge graph system based on the association relationship between elements, and reflect the association of knowledge systems and elements in various fields in the knowledge graph; Step 6, use a graph database for storage. 2.The multi-domain knowledge fusion method based on semantic tree according to claim 1, characterized in that, In step 2, the association rule is used to reflect the interdependence and association between one thing and other things, and is used to mine the correlation between valuable data items from a large amount of data; The association strength of the association rule is described by support, confidence and lift; The extraction process of the association rule is to find all association rules with support >= minimum support and confidence >= minimum confidence; The association rules that meet the conditions can be mined in two steps: first, generate frequent item sets, that is, find all item sets that meet the minimum support, and then generate rules that meet the minimum confidence based on the frequent item sets; Through association rule mining, the correlation between different field knowledge is obtained, which directly reflects the overall correlation degree of the multi-field knowledge system and lays a theoretical foundation for the establishment of a comprehensive knowledge graph. 3.The multi-domain knowledge fusion method based on semantic tree according to claim 1, characterized in that, In step 4, the most authoritative node is determined by mutual voting method, if the data volume of the knowledge graph is very large, the block method is used to realize it, first calculate the PageRank value of each block graph, according to the correlation between data blocks, get the new graph PageRank value, and then iterate repeatedly to analyze the most authoritative node; A multi-feature method based on node attributes and node relationships is used to analyze the authoritative nodes by comprehensively analyzing the node attributes and relationships.

Images