Abstract
This paper describes our participating system run to the argumentative text understanding shared task for AI Debater at NLPCC 2021 (http://www.fudan-disc.com/sharedtask/AIDebater21/tracks.html). The tasks are motivated towards developing an autonomous debating system.
We make an initial attempt with Track-3, namely, argument pair extraction from peer review and rebuttal where we extract arguments from peer reviews and their corresponding rebuttals from author responses. Compared to the multi-task baseline by the organizers, we introduce two significant changes: (i) we use ERNIE 2.0 token embedding, which can better capture lexical, syntactic, and semantic aspects of information in the training data, (ii) we perform double attention learning to capture long-term dependencies.
Our proposed model achieves the state-of-the-art results with a relative improvement of 8.81% in terms of F1 score over the baseline model. We make our code available publicly at https://github.com/guneetsk99/ArgumentMining_Shared