面向知识图谱的自然语言处理——知识图谱原理与应用-3

背景知识

image-20250530102706259

image-20250530103041006

image-20250530103106083

Information Retrieval-based

image-20250530103759411

Question Answering with Subgraph Embeddings

image-20250530103945733

1.对用户输入的自然语言进行word embedding，得到自然语言的embedding
2.对输入的问题中的topic entity链接到图谱中某个具体的实体，对映射的实体找到k步内的邻居，找到候选节点
3.对相应实体的k步内的subgraph做embedding
4.算两个embedding之间的得分

image-20250530105250161

问题侧的表征方法

1

答案侧的表征方法

Single Entity

image-20250530105815795

Path Representation

image-20250530105849144

image-20250530105911606

Subgraph Representation

image-20250530110210152

屏幕截图 2025-05-30 110217

得分函数

image-20250530110325927

Question Answering over Freebase with Multi-Column Convolutional Neural Networks

image-20250530110958229

image-20250530111026828

image-20250530111225886

image-20250530111319878

image-20250530111401323

image-20250530111414580

image-20250530111433710

Semantic Parsing

image-20250530113844920

image-20250530114058877

Semantic Parsing via Staged Query Graph Generation

image-20250530115301256

image-20250530115328171

image-20250530115238464

image-20250530115447428

Challenges of Complex Questions

image-20250530115924456

image-20250530115947743

image-20250530120006506

image-20250530120020922

Our Approach——Data Driven & Relation-first framework gAnswer

image-20250530120319721

image-20250530120412633

image-20250530141327582

image-20250530141429027

保留歧义(Ambiguity)，Paul Anderson可能对应多个实体，不进行歧义消除

image-20250530141805153

image-20250530141947380

通过子图匹配的方法，可以消除歧义，同时也可以找到答案

Relation First

image-20250530142525286

抽取关系，再将孤立的关系拼接

image-20250530142647575

image-20250530142829611

Limitations

10

Node First

image-20250530143356441

image-20250530143446567

11

映射到知识图谱当中查看两者最有可能的关系是什么

12

posted @ 2025-05-30 14:44  狐狸胡兔  阅读(10)  评论(0)    收藏  举报