US 12,333,447 B2
Path optimization using multi-domain knowledge graphs for inference in dialog systems
Shinya Taguchi, Tokyo (JP); and Yusuke Seto, Tokyo (JP)
Assigned to MITSUBISHI ELECTRIC CORPORATION, Tokyo (JP)
Filed by Mitsubishi Electric Corporation, Tokyo (JP)
Filed on Feb. 9, 2021, as Appl. No. 17/171,058.
Application 17/171,058 is a continuation of application No. PCT/JP2018/036198, filed on Sep. 28, 2018.
Prior Publication US 2021/0166141 A1, Jun. 3, 2021
Int. Cl. G06N 5/04 (2023.01); G06F 16/901 (2019.01); G06F 17/18 (2006.01); G06N 5/02 (2023.01)
CPC G06N 5/04 (2013.01) [G06F 16/9024 (2019.01); G06F 17/18 (2013.01); G06N 5/02 (2013.01)] 8 Claims
OG exemplary drawing
 
1. An inference device comprising:
processing circuitry to generate integrated information by combining items of information respectively belonging to domains different from each other included in dynamically changing external information via forward chaining by using knowledge information including information regarding a human's condition and information regarding a human's action and provided from a knowledge base and an inference rule provided from a rule database; and
an operation history storage that stores operation history records of a user, wherein
each of the items of information respectively belonging to domains different from each other is information that can be represented as a directed graph including a node and an edge,
the integrated information is an integrated graph generated by combining directed graphs that are the items of information respectively belonging to different domains,
the processing circuitry calculates an importance level of a node as a component of the integrated graph from a probability of arriving at the node in a stationary state reached by performing a random walk on the integrated graph or by using an algorithm of PageRank, and
the processing circuitry
sets a node that is a component of the directed graph and serves as a starting point of an inference according to the inference rule, as an observation node,
generates an observation vector in which a weight of the observation node is set at a value higher than or equal to 0 and weights of nodes other than the observation node are set at 0,
calculates the importance level based on the observation vector,
updates the observation vector based on the operation history records, and
calculates the importance level based on the updated observation vector.