公开(公告)号：US11562133B2

公开(公告)日：2023-01-24

申请号：US16696476

申请日：2019-11-26

Applicant: FOUNDATION OF SOONGSIL UNIVERSITY-INDUSTRY COOPERATION

Inventor： Young Tack Park ,  Wan Gon Lee ,  Jagvaral Batselem ,  Hyun Young Choi ,  Ji Houn Hong

IPC: G06F40/20 ,  G06F40/166 ,  G06K9/62 ,  G06N20/20

Abstract: Provided is an incorrect triple detection system including a triple selector configured to select a target triple (subject, type, object) in a knowledge base, a sampler configured to create a sentence model by connecting object triples sharing entities included in the target triple, a model builder configured to embed the sentence model into a vector space to create a training entity vector and build an embedding model, and an incorrect triple detector configured to detect an incorrect triple by inputting a test triple into the embedding model.

公开(公告)号：US11562177B2

公开(公告)日：2023-01-24

申请号：US16697238

申请日：2019-11-27

Applicant: Foundation of Soongsil University-Industry Cooperation

Inventor： Young Tack Park ,  Jagvaral Batselem ,  Wan Gon Lee

IPC: G06K9/00 ,  G06K9/62 ,  G06N3/08 ,  G06N3/04 ,  G06F17/16 ,  G06N20/20 ,  G06N20/10 ,  G06N5/04

Abstract: A triple verification method is provided. The triple verification method includes setting a triple having a source entity, a target entity, and a relation value between the source entity and the target entity by a setting unit, extracting a plurality of intermediate entities associated with the source entity and the target entity by the setting unit, defining a connection relation between the intermediate entity, the source entity, and the target entity and generating a plurality of connection paths connecting the source entity, the intermediate entity, and the target entity by a path generation unit, generating a matrix by embedding the plurality of connection paths into vector values by a first processing unit, calculating a feature map by performing a convolution operation on the matrix by a second processing unit, generating an encoding vector for each connection path by encoding the feature map by applying a bidirectional long short-term memory neural network (BiLSTM) technique by a third processing unit, and generating a state vector by summing the encoding vectors for each connection path by applying an attention mechanism and verifying the triple based on a similarity value between the relation value of the triple and the state vector by a determination unit.