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](https://ayimd-pic.oss-cn-guangzhou.aliyuncs.com/image-20230621153922205.png)
while the hypergraph convolutional networks is:
![image-20230621155931211](https://ayimd-pic.oss-cn-guangzhou.aliyuncs.com/image-20230621155931211.png)
DHGCN
the directed hypergraph can be denoted by two incidence matrices \(H^{head}\) and \(H^{tail}\)
the degree:
![image-20230621160733846](https://ayimd-pic.oss-cn-guangzhou.aliyuncs.com/image-20230621160733846.png)
the directed hypergraph convolutional networks is:
![image-20230621160751671](https://ayimd-pic.oss-cn-guangzhou.aliyuncs.com/image-20230621160751671.png)
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
- obtain seed entities from the question by entity linking
- get the entities set within L hops to form a subgraph
- get \(H^{head}\) and \(H^{tail}\)
Input Encoder
apply a bi-LSTM to encode question and obtain hidden states \(H\in R^{|q|\times h}\),we assume h=d
employ co-attention to learn query-aware entity representation
Reasoning over Hypergraph
![image-20230621163315446](https://ayimd-pic.oss-cn-guangzhou.aliyuncs.com/image-20230621163315446.png)
![image-20230621172646592](https://ayimd-pic.oss-cn-guangzhou.aliyuncs.com/image-20230621172646592.png)
Learn Relation Representation Explicitly
combine entity embedding and co-attention
propagation
aggregation
Allocate Relation Weights Dynamically(dynamically allocated hop-by-hop)
use co-attention to cal \(R_{co\_attn}\)
compute the weight of edge
Update Entity Adaptively