nanguoshun / LSR

Pytorch Implementation of our ACL 2020 Paper "Reasoning with Latent Structure Refinement for Document-Level Relation Extraction"
127 stars 21 forks source link

Why use matrix tree + dense gcn #40

Open chijames opened 3 years ago

chijames commented 3 years ago

Hi,

By using the matrix tree theorem, you are basically assuming all the node entities form a dependency tree. However, a dense gcn assumes all entities form a complete graph. These two assumptions seem to be contradictory to me. Also, by looking at the right-most column of figure 5, the induced structures are also not trees.

Please correct me if I misunderstand something. Thanks!