Methods for building and updating knowledge graphs

Abstract

A method for establishing and updating knowledge graphs uses the graph relationships between multiple image knowledge graphs as the "cause" and the updating of the image node relationship network as the "effect." Therefore, this invention, through causal relationships, allows multiple image knowledge graphs to achieve a process of refining and eliminating unnecessary information, which is beneficial for artificial intelligence algorithms to understand things through the hierarchical knowledge graph updated by this invention. Furthermore, since the morphological characteristics of objects change with time and location, this invention can also define the relationship between two objects through a text knowledge graph, and then use this relationship to update the node layer of the image knowledge graph, further facilitating artificial intelligence understanding.

Claims

1. A method for establishing and updating a knowledge graph, executed on a processor, comprising: reading a first image knowledge graph and a second image knowledge graph from a storage device, wherein each of the first and second image knowledge graphs includes an image node, multiple sub-image nodes connecting the root image node, and multiple grandchild image nodes connecting each sub-image node; reading a graph relationship index from the storage device, the graph relationship index representing whether the first and second image knowledge graphs are of the same or different types; executing an image similarity recognition model to identify image similarity values ​​of multiple items of two corresponding sub-image nodes and multiple items of two corresponding grandchild image nodes in the first and second image knowledge graphs, and storing them in the storage device; determining whether the image similarity values ​​of the two corresponding sub-image nodes and the image similarity values ​​of the two corresponding grandchild image nodes are greater than or equal to an image threshold to generate a determination result and store it in the storage device; Based on the judgment result, an image node connection index between two corresponding child image nodes and between two corresponding grandchild image nodes of the first image knowledge graph and the second image knowledge graph are established through a decision tree judgment process, and stored in the memory. The image node connection indices represent identical, equivalent, related, or zero-related relationships. When the graph relationship is of the same kind, the following steps are performed to update the relationship network of the first image knowledge graph and the second image knowledge graph and store it in the storage: merging two corresponding child image nodes and two corresponding grandchild image nodes whose image node connection indices are identical or equivalent through a node merging instruction; deleting related image node connection indices from the storage; adding two corresponding grandchild image nodes whose image node connection indices are zero-related to the child image nodes of each grandchild image node; and adding two corresponding child image nodes whose image node connection indices are zero-related to the root image node of each child image node. When the graph relationship is of different kinds, the following steps are performed to update the relationship network of the first image knowledge graph and the second image knowledge graph and store it in the storage: retaining the image node connection indices that are identical in the storage; and deleting image node connection indices that are equivalent, related, or zero-related from the storage. The image similarity values ​​of the multiple items include similarity values ​​of image information for a type item, a size item, a texture item, and a color item. The decision tree judgment process proceeds in the following order: type item, size item, texture item, and color item. When the processor determines that the type item, size item, texture item, and color item are similar, the established image node connection index is the same. When the processor determines that the type item, size item, texture item, and color item are similar, the established image node connection index is equivalent. When the processor determines that the type item, size item, texture item, and color item are similar or dissimilar, the established image node connection index is related. When the processor determines that the type item, size item, texture item, and color item are similar or dissimilar, the established image node connection index is equivalent. When the processor determines that the type item is similar, the size item is dissimilar, the texture item is dissimilar, and the color item is similar or dissimilar, the established image node connection index is correlated;When the processor determines that the type item is dissimilar, the size item is dissimilar, the texture item is dissimilar, and the color item is similar, the established image node connection index is correlated; when the processor determines that the type item is dissimilar, the size item is dissimilar, the texture item is dissimilar, and the color item is dissimilar, the established image node connection index is zero correlated.

2. A method for establishing and updating relationships in a knowledge graph, executed on a processor, comprising: reading a first text knowledge graph and a second text knowledge graph from a storage device, wherein each of the first and second text knowledge graphs comprises a root node, one or more child nodes connecting the root node, one or more grandchild nodes connecting each child node, and one or more feature text nodes connecting each grandchild node, wherein the child nodes comprise weights; executing a text similarity recognition model to identify and store the text similarity values ​​of two corresponding feature text nodes in the first and second text knowledge graphs in the storage device; determining whether the text similarity values ​​of two corresponding feature text nodes in the first and second text knowledge graphs are greater than or equal to a text threshold, generating a determination result and storing it in the storage device; and establishing a text node connection index between two corresponding feature text nodes in the first and second text knowledge graphs based on the determination result and storing it in the storage device. The connection index of a grandchild node between two corresponding grandchild nodes of the first text knowledge graph and the second text knowledge graph is determined based on the text node connection index and stored in the storage. The connection index of the grandchild nodes under the child node with the highest weight is updated based on the grandchild node connection index of the child node with the lowest weight. Feature text nodes are added based on the updated grandchild node connection index, forming an updated text knowledge graph and stored in the storage. A graph relationship index is defined based on the updated text knowledge graph and stored in the storage, where the graph relationship index represents the same or different types. A first image knowledge graph and a second image knowledge graph are read from the storage, each containing an image node, multiple child image nodes connecting the root image node, and multiple grandchild image nodes connecting each child image node. The graph relationship index is read from the storage, where the graph relationship index represents whether the first image knowledge graph and the second image knowledge graph are the same or different types. An image similarity recognition model is executed to identify the image similarity values ​​of multiple items between two corresponding sub-image nodes and multiple items between two corresponding grandchild image nodes in the first and second image knowledge graphs, and these values ​​are stored in the storage. A judgment result is generated and stored in the storage, determining whether the image similarity values ​​of the two corresponding sub-image nodes and the two corresponding grandchild image nodes are greater than or equal to an image threshold. Based on the judgment result, an image node connection index between two corresponding sub-image nodes and between two corresponding grandchild image nodes is established through a decision tree judgment process, and this index is stored in the storage. The image node connection indices represent identical, equivalent, related, or zero-related relationships. When the graph relationship is of the same kind, the following steps are performed to update the relationship network of the first image knowledge graph and the second image knowledge graph and store it in the storage: merging two corresponding child image nodes and two corresponding grandchild image nodes whose image node connection indices are identical or equivalent through a node merging instruction; deleting related image node connection indices from the storage; adding two corresponding grandchild image nodes whose image node connection indices are zero-related to the child image nodes of each grandchild image node; and adding two corresponding child image nodes whose image node connection indices are zero-related to the root image node of each child image node. When the graph relationship is of different kinds, the following steps are performed to update the relationship network of the first image knowledge graph and the second image knowledge graph and store it in the storage: retaining the image node connection indices that are identical in the storage; and deleting image node connection indices that are equivalent, related, or zero-related from the storage.

3. The method for establishing and updating the knowledge graph as described in request item 2, wherein, The image similarity value of the multiple items includes the similarity value of image information of a shape item, a size item, a texture item, and a color item.

4. The method for establishing and updating the knowledge graph as described in request item 3, wherein, The decision tree's judgment process proceeds in the following order: type item, size item, texture item, and color item. When the processor determines that the type item, size item, texture item, and color item are similar, the established image node connection index is "same." When the processor determines that the type item, size item, texture item, and color item are similar, the established image node connection index is "equivalent." When the processor determines that the type item, size item, texture item, and color item are similar or dissimilar, the established image node connection index is "correlated." When the processor determines that the type item, size item, texture item, and color item are similar or dissimilar, the established image node connection index is "equivalent." When the processor determines that the type item, size item, texture item, and color item are similar or dissimilar, the established image node connection index is "correlated." When the processor determines that the type item is dissimilar, the size item is similar, the texture item is similar, and the color item is similar or dissimilar, the established image node connection index is correlated; when the processor determines that the type item is dissimilar, the size item is similar, the texture item is dissimilar, and the color item is similar or dissimilar, the established image node connection index is correlated; when the processor determines that the type item is dissimilar, the size item is dissimilar, the texture item is similar, and the color item is similar or dissimilar, the established image node connection index is correlated; when the processor determines that the type item is dissimilar, the size item is dissimilar, the texture item is dissimilar, and the color item is similar, the established image node connection index is correlated; when the processor determines that the type item is dissimilar, the size item is dissimilar, the texture item is dissimilar, and the color item is dissimilar, the established image node connection index is zero-correlation.

5. The method for establishing and updating the knowledge graph as described in request item 2, wherein, When the processor determines that all feature text nodes under the two corresponding grandchild nodes of the first and second text knowledge graphs have the text node connection index set, the processor establishes that the grandchild node connection indexes of the two corresponding grandchild nodes are completely identical and stores them in the storage; when the processor determines that only some feature text nodes under the two corresponding grandchild nodes of the first and second text knowledge graphs have the text node connection index set, the processor establishes that the grandchild node connection indexes of the two corresponding grandchild nodes are not completely identical and stores them in the storage; when the processor determines that none of the feature text nodes under the two corresponding grandchild nodes of the first and second text knowledge graphs have the text node connection index set, the processor establishes that the grandchild node connection indexes of the two corresponding grandchild nodes are different and stores them in the storage.

6. The method for establishing and updating the knowledge graph as described in request item 5, wherein, When the processor determines that the connection metrics of the grandchild nodes under the child node with the highest weight are completely identical, and the connection metrics of the grandchild nodes under the child node with the lower weight are not completely identical or different, the processor modifies the connection metrics of the grandchild nodes under the child node with the lower weight from not completely identical or different to completely identical and stores it in the memory; when the processor determines that the connection metrics of the grandchild nodes under the child node with the highest weight are not completely identical or different, the processor does not modify the connection metrics of the grandchild nodes under the child node with the lower weight.

Images