Document-level Relation Extraction as Semantic Segmentation

Introduction

• What is the name of the DocuNET paper?

  Document-level Relation Extraction as Semantic Segmentation

• What are the main contributions of the DocuNET paper?
  • treat document level relation extraction as a semantic segmentation task
  • take advantage of global dependency among relational triples
  • leverage U-NET inspired semantic segmentation architecture

• Document level relation extraction extracts relation information from multiple sentences

  

• What are some difficulties with document level relation extraction?
  • subject and object entities may not appear in the same sentence
• Previous approaches for document level RE include graph based and transformer model based

Method

• DocuNET leverages a entity level relation matrix
  • each cell is a relation type

  

• DocuNET problem statement: output matrix is n x n giving each entity relation

• DocuNET encodes a document using a transformer with special entity mention boundary tokens

• DocuNET embeds longer documents using dynamic window pooling
  • average embeddings of overlapping tokens of different windows\
• DocuNET entity-entity relevance vector from similarity method is concatenation of element wise similarity, cosine similarity and bilinear similarity
  • based only on embedding for each entity
  • alternative to a content based method

  

• DocuNET content based strategy leveraging entity based attention and the global document embedding

  

• DocuNET UNET architecture

  

Results

• DocuNET case study

  

• DocuNET strong results on Biomedical datasets

  

• DocuNET results on DocRED

  

Reference

@misc{https://doi.org/10.48550/arxiv.2106.03618,
  doi = {10.48550/ARXIV.2106.03618},
  
  url = {https://arxiv.org/abs/2106.03618},
  
  author = {Zhang, Ningyu and Chen, Xiang and Xie, Xin and Deng, Shumin and Tan, Chuanqi and Chen, Mosha and Huang, Fei and Si, Luo and Chen, Huajun},
  
  keywords = {Computation and Language (cs.CL), FOS: Computer and information sciences, FOS: Computer and information sciences},
  
  title = {Document-level Relation Extraction as Semantic Segmentation},
  
  publisher = {arXiv},
  
  year = {2021},
  
  copyright = {arXiv.org perpetual, non-exclusive license}
}