Prompt for Extraction? PAIE: Prompting Argument Interaction for

Introduction

• What is the name of the PAIE paper?

  Prompt for Extraction? PAIE: Prompting Argument Interaction for Event Argument Extraction

• What are the main contributions of the PAIE paper?
  • prompt tuning method for EAE
• Two main approaches for EAE include semantic role labeling and question answering/text generation
  • Semantic Role Labeling: Identify spans + classify their roles

• Semantic Role labeling approaches for EAE have two basic steps: identifying candidate spans and classifying their roles

• One disadvantage of QA based models for EAE is the need to frame a separate question for each argument
  • ex: NLI+ paper need to use multiple templates and forward passes through NLI model
  • Generative methods can generate all at once:
    • however current performance of these types of methods are not great
• What are the main goals of the PAIE paper?
  • extract all arguments efficiently
  • how to capture argument interactions in long text on the fly
  • how can we leverage knowledge from
• EAE tasks can be framed on both the sentence and document level
  • some may require identifying antecedents of different arguments

  

• QA-based methods usually adopt the thresholding strategy which must be manually tuned
  • likelihood threshold for QA argument extraction answer

Method

• The three main steps of the PAIE method are 1) prompt creation, 2) span selector decoding and 3) span prediction
  • prompt is generated from a series of queries
  • prompted selector decoding: spans selected based on input, prompt and query
  • prompted span selection: selectors for start and end of optimal span
• PAIE feeds span and context into BART backbone
  • include special markers before and after trigger tokens

  

• PAIE span selector head uses start and end token hidden representation as features
  • can simultaneously score all roles by combing all k queries in one matrix multiply by the meta prediction head

  

Results

• PAIE main results on ACE and Wikievents

  

Conclusions

• Does worse than NLI+ paper on wikievents but seemingly better on ACE

Reference

@misc{https://doi.org/10.48550/arxiv.2202.12109,
  doi = {10.48550/ARXIV.2202.12109},
  
  url = {https://arxiv.org/abs/2202.12109},
  
  author = {Ma, Yubo and Wang, Zehao and Cao, Yixin and Li, Mukai and Chen, Meiqi and Wang, Kun and Shao, Jing},
  
  keywords = {Computation and Language (cs.CL), Artificial Intelligence (cs.AI), FOS: Computer and information sciences, FOS: Computer and information sciences},
  
  title = {Prompt for Extraction? PAIE: Prompting Argument Interaction for Event Argument Extraction},
  
  publisher = {arXiv},
  
  year = {2022},
  
  copyright = {Creative Commons Attribution 4.0 International}
}