Efficient Argument Structure Extraction with Transfer Learning and Active Learning

This repository contains AMPERE++, the dataset associated with the paper “Efficient Argument Structure Extraction with Transfer Learning and Active Learning. Xinyu Hua and Lu Wang, Findings of ACL 2022”.

It contains 400 academic peer reviews of ICLR 2018, collected from openreview.net in the prior work “Argument Mining for Understanding Peer Reviews”. Each review is already segmented into propositions on sub-sentence level. We further annotate support/attack relations among these propositions.

For experiment purposes, we adopt the original train/val/test split. We provide the jsonl format with the following fields:

• doc_id: a unique review document id
• text: the list of propositions
• relations: the list of relations, each relation consists of the following:
  • head: the index of head proposition (starting from 0)
  • tail: the index of tail proposition (starting from 0)
  • type: the relation type, either `support` or `attack`.