US 12,242,975 B2
Querying knowledge graphs with sub-graph matching networks
Lingfei Wu, Elmsford, NY (US); and Chen Wang, Chappaqua, NY (US)
Assigned to INTERNATIONAL BUSINESS MACHINES CORPORATION, Armonk, NY (US)
Filed by International Business Machines Corporation, Armonk, NY (US)
Filed on Oct. 1, 2020, as Appl. No. 17/061,011.
Prior Publication US 2022/0108188 A1, Apr. 7, 2022
Int. Cl. G06N 5/02 (2023.01); G06F 17/16 (2006.01); G06N 3/045 (2023.01); G06N 5/04 (2023.01)
CPC G06N 5/02 (2013.01) [G06F 17/16 (2013.01); G06N 3/045 (2023.01); G06N 5/04 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A system, comprising:
a memory that stores computer executable components; and
a processor, operably coupled to the memory, and that executes the computer executable components stored in the memory, wherein the computer executable components comprise:
a question answering over knowledge graph component that:
trains a bidirectional gated graph sequence neural network to encode first neural network embeddings of graph structure information of knowledge graph subgraphs;
trains a bidirectional message passing neural network to encode second neural network embeddings of graph structure information of question graphs;
encodes first graph structure information of a knowledge graph subgraph of a knowledge graph into a first neural network embedding using the bidirectional gated graph sequence neural network, wherein the encoding the first graph structure information comprises generating, for a node of the knowledge graph subgraph:
a node vector representation for the node of the knowledge graph subgraph based on node vector representations for one or more neighboring nodes of the knowledge graph subgraph to the node in the knowledge graph subgraph, and edge representations for one or more edges between the node of the knowledge graph subgraph and the one or more neighboring nodes of the knowledge graph subgraph; and
encodes second graph structure information of a question graph into a second neural network embedding using the bidirectional message passing neural network, wherein the encoding the second graph structure information comprises generating, for a node of the question graph:
a vector representation for the node of the question graph at a layer of the question graph based on vector representations for one or more neighboring nodes of the question graph to the node of the question graph at another layer of the question graph; and
a match component that executes a similarity algorithm to compute a matching score from the first neural network embedding and the second neural network embedding, wherein the matching score characterizes an amount of similarity between the question graph and the knowledge graph subgraph.