DHGCN(2HRDR)

propose DHGCN,2HRDR

Background

task: 基于知识图谱的问答系统，在知识图谱中检索与问题相关的多个元组

contribution: propose a convolutional network for directed hypergraph

DHGCN

HGCN

given a hypergraph \(G=(V,E,W)\), as well as the incidence matrix \(H\in R^{|V|\times |E|}\)

the edge and vertex degrees:

image-20230621153922205

while the hypergraph convolutional networks is:

image-20230621155931211

IMG_0183(20230621-201655)

DHGCN

the directed hypergraph can be denoted by two incidence matrices \(H^{head}\) and \(H^{tail}\)

the degree:

image-20230621160733846

the directed hypergraph convolutional networks is:

image-20230621160751671

2HRDR

Task Definition

given a knowledge graph \(K=(V,E,T)\) and \(q=(w_1,w_2,\cdots ,w_{|q|})\).

the task aims to pick the answers from V.

Method

Directed Hypergraph Retrieval and Construction

find subgraph

1. obtain seed entities from the question by entity linking
2. get the entities set within L hops to form a subgraph
3. get \(H^{head}\) and \(H^{tail}\)

Input Encoder

1. apply a bi-LSTM to encode question and obtain hidden states \(H\in R^{|q|\times h}\),we assume h=d

2. employ co-attention to learn query-aware entity representation

   

   image-20230621170223607

Reasoning over Hypergraph

image-20230621163315446

image-20230621172646592

1. Learn Relation Representation Explicitly

   1. combine entity embedding and co-attention

      

      image-20230621173702999

   2. propagation

      

      image-20230621173728920

   3. aggregation

      

      image-20230621173815550

2. Allocate Relation Weights Dynamically(dynamically allocated hop-by-hop)

   1. use co-attention to cal \(R_{co\_attn}\)

   2. compute the weight of edge

      

      image-20230621174053983

      

      image-20230621174101797

3. Update Entity Adaptively

   

   image-20230621174229381