Relation Extraction

1. Matching the Blanks: Distributional Similarity for Relation Learning [ACL 2019] Livio Baldini Soares, Nicholas FitzGerald, Jeffrey Ling, Tom Kwiatkowski.
   • How to solve relation extraction task by using BERT
   • three representation
     1. standard BERT
     2. add Position embed
     3. Entity Mark
   • three output
     1. CLS output
     2. Mention level pooling
     3. Mention start
   • random replace the Entity -> [BLANK]
   • negative sample
2. Inducing Relational Knowledge from BERT [AAAI 2020] Zied Bouraoui, Jose Camacho-Collados, Steven Schockaert.
   • using the middle sentence to represent the relation of the pair.
   • choose one temperate to do the cloze task.
3. Entity-Relation Extraction as Multi-Turn Question Answering [ACL 2019] Xiaoya Li, Fan Yin, Zijun Sun, Xiayu Li, Arianna Yuan, Duo Chai, Mingxin Zhou, Jiwei Li.
   • Start-end + Multi-time => RL. The model also a pipeline model.
4. Twenty-five years of information extraction [NLE 2019] Ralph Grishman
   • From Three Corpus to introduce the problem and method in Information Extraction IE in past 25 years.
5. Improving Entity Linking by Modeling Latent Entity Type Information [AAAI 2020] Shuang Chen, Jinpeng Wang, Feng Jiang, Chin-Yew Lin.
   • Improve the embedding by [MASK]
6. Relation of the Relations: A New Paradigm of the Relation Extraction Problem [-] Zhijing Jin, Yongyi Yang, Xipeng Qiu, Zheng Zhang.
   • Focus on Relation of Relations, directly learn the entity pairs matrix instead of one by one.
   • Statistic the co-occurrence of relation pairs.
   • BioROR. -> GNN.
   • MultiRoR. -> Transformer.
7. Relation Extraction using Explicit Context Conditioning [NAACL 2019] Gaurav Singh, Parminder Bhatia.
   • Motivation:
     1. Get two-hop link relation.
     2. Get long dependencies.
   • Insight:
     1. Build first-order relation scores based on bi-affine transformer.
     2. Build second-order relation scores based on conditional score (combine two first-order scores which have same item k). => This part should improve by using GCN or other graph embedding architecture.
     3. One efficient implement which reduce the second order time cost.