Graph-Based Reasoning over Heterogeneous External Knowledge for Commonsense Question Answering

论文目的

实现常识问答——常识问答要用到外部知识，现有的常识问答方法有两种：they either generate evidence from human-annotated evidence or extract evidence from a homogeneous knowledge source like structured knowledge ConceptNet. 但是they fail to take advantages of both knowledge sources simultaneously.

论文方法

Our approach consists of two parts: knowledge extraction and graph-based reasoning

knowledge extraction

原文：extract knowledge from structured knowledge base ConcpetNet and Wikipedia plain texts according to the given question and choices.We construct graphs to utilize the relational structures of both sources.

从conceptnet中抽取知识

对于每个问题和选择，我们在conceptnet图中定义它们的实体。然后从问题实体到选择实体找出至少三条路，然后构成graphs

从Wikipedia中抽取知识

先把问题和选择中的停用词去除，拼接后查询，搜索前K名的作为证据，证据按照eSemantic Role Labeling (SRL)方法结构化构图。

graph-based reasoning

原文: we propose two graph-based modules which consists of a graph-based contextual word representation learning module and a graph-based inference module to infer final answers

graph-based contextual representation learning module

借助图信息学习文本word表示重新定义word距离

graph-based inference module

通过图卷积网络，利用邻居信息获得节点表示，并且汇总图表示推测结果

实验方法

baseline

For the compared methods, we select models and classify them into 4 groups.
Group 1: models without descriptions or papers,
Group 2: models without extracted knowledge,
Group 3: models with extracted structured knowledge and
Group 4: models with extracted unstructured knowledge.

实验内容

1. 在CommonsenseQA上对比baseline效果
2. Ablation Study——对比模型简化版本以证明每个模块的有效性
3. case study
4. 误差分析——the lack of evidence, similar evidence and dataset noise

论文亮点

实验做得极其充分