Improving Multi-hop Question Answering over Knowledge Graphs using Knowledge Base Embeddings

0.Summary

 

1.Reseach Objective

this research on multi-hop KGQA( Question Answering over Knowledge Graphs) has attempted to handle KG sparsity using 

 knowledge enbedding  predict missing link on KG and Improve the accuracy of multi-hop KGQA。

EmbedKGQA whilch is proposed on the paper is particu-larly effective in performing multi-hop KGQAover sparse KGs.

2.Background and Problem

In prior work (Li et al., 2018) TransE,(Bordes et al., 2013) embeddings have been used toanswer factoid based questions(基于事实问题) but it does not work for multi-hop question answering.

• Two approaches existed
  • impose heuristic neighborhood limits(启发式限制搜索，不一定能找到答案)
  • used an external text corpus to handle KG sparsity(使用额外的问题处理KG的稀疏性)

problem : handle the KG sparsity to improve the performence of multi-hop question answer.

3.Method

example:

 

 

model:

•  KG Embedding Module
  • use complEx Embeddings to create vector
  • use entity embeddings to learn a triple scoring function  between head entity and question, and answer entity.
• Question Embedding Module
  • the questionqusing RoBERTa (Liuet al., 2019) into a 768-dimensional vector
  • passed through 4 fully connected linear layerswith ReLU activation and finally projected onto thecomplex space .
• Answer Selection Module
  • selectthe entity with the highest score
  • relation matching

4.Evaluation

• A novel method for complete knowledge graph is proposed

• The details are incomplete

• use mateQa and MetaQa (全部和百分之50数据集进行比较，指标是答案和推理是否相似)

      

 

5.Conclusion

 we propose EmbedKGQA, a novelmethod for Multi-hop KGQA. KGs are often in-complete and sparse which poses additional challenges for multi-hop KGQA methods.

KG Embedding suggest the linkprediction properties of KG embeddings can be uti-lized to mitigate the KG incompleteness problemin Multi-hop KGQA.

6.contributions

• We propose EmbedKGQA, a novel methodfor the multi-hop KGQA task. To the bestof our knowledge, EmbedKGQA is the firstmethod to use KG embeddings for this task.EmbedKGQA is particularly effective in per-forming multi-hop KGQA over sparse KGs.
• EmbedKGQA relaxes the requirement of an-swer selection from a pre-specified localneighborhood, an undesirable constraint im-posed by previous methods for this task.
• EmbedKGQA relaxes the requirement of an-swer selection from a pre-specified localneighborhood, an undesirable constraint im-posed by previous methods for this task.

7. Reference

 

8.Notes