REM-Net: Recursive Erasure Memory Network for Explanation Refinement on Commonsense Question Answering

EasyChair Preprint 4973

Abstract

Commonsense Question Answering (commonsense QA) tasks aim to examine QA systems' capability of reasoning over context with complicated logical relationships and implicit commonsense knowledge. A desirable model should be able to provide not only correct answers but also persuasive explanations. Current works incorporate external knowledge presuming that valid explanations are included. However, the explanations are usually confounded and need further distinguishment. In this work, we propose a recursive erasure memory network (REM-Net), which learns to refine explanations for more precise interpretation while reasoning to obtain correct answers. REM-Net integrates a pre-trained knowledge graph generator, to provide possible explanations based on the commonsense question, and a recursive erasure memory module (REM), which refines the explanations. The REM module recursively erases confounding explanations to ensure that the model captures the most crucial clues. Experimental results on multiple commonsense QA benchmarks demonstrate that our REM-Net outperforms the competing methods. The case study also shows the model's ability to find more precise explanations.

Keyphrases: Commonsense Question Answering, Recursive Erasure Memory Network, explanation refinement

@booklet{EasyChair:4973,
  author    = {Yinya Huang and Meng Fang and Xunlin Zhan and Qingxing Cao and Xiaodan Liang and Liang Lin},
  title     = {REM-Net: Recursive Erasure Memory Network for Explanation Refinement on Commonsense Question Answering},
  howpublished = {EasyChair Preprint 4973},
  year      = {EasyChair, 2021}}