Modeling Semantics with Gated Graph Neural Networks for Knowledge Base Question Answering笔记
来源: COLING 2018 Long Paper
原文
Movation
以往的模型在复杂问题(问题实体和答案实体之间相隔较远)上的表现很差。
作者认为:
We claim that one needs to explicitly model the semantic structure to be able to find the correct semantic parse for complex questions.
这里所说的“semantic parse(semantic graph)”如图1所示,之后可以转化为查询从KB中得到答案:
Semantic parsing
Semantic graphs
用图的方式对问题进行结构化表示,
Our semantic graphs consist of a question variable node (q), Wikidata entities ( Taylor Swift ), relation types from Wikidata ( PERFORMER ) and constraints (see Figure 3 for an example graph with a constraint).
Semantic graph construction
可能借鉴了强化学习的思路。
state: . 表示当前的图, 是在问题中出现但还没有加入图的实体。显然,
action: .
Representation learning
Deep Convolutional Networks: 使用DCNN得到向量 作为问题的表达。
结论与思考
旧问题,旧思路,新方法,扩展了查询图的定义,并且first to use GGNNs for semantic parsing and KB QA。