基于推理链的多跳问答对抗攻击和对抗增强训练方法(Reasoning Chain Based Adversarial Attack and Adversarial Augmentation Training for Multi-hop Question Answering)
Jiayu Ding (佳玙丁,), Siyuan Wang (王思远), Zhongyu Wei (魏忠钰), Qin Chen (陈琴), Xuanjing Huang (黄萱菁)
Abstract
“本文提出了一种基于多跳推理链的对抗攻击方法,通过向输入文本中加入对抗性的攻击文本,并测试问答模型在干扰数据下生成答案的准确性,以检测问答模型真正执行多跳推理的能力和可解释性。该方法首先从输入文本中抽取从问题实体到答案实体的推理链,并基于推理链的特征把多跳问题分为了不同的推理类型,提出了一个模型来自动化实现问题拆解和推理类型预测,然后根据推理类型对原问题进行修改来构造攻击干扰句。实验对多个多跳问答模型进行了对抗攻击测试,所有模型的性能都显著下降,验证了该攻击方法的有效性以及目前问答模型存在的不足;向原训练集中加入对抗样本进行增强训练后,模型性能均有所回升,证明了本对抗增强训练方法可以提升模型的鲁棒性。”- Anthology ID:
- 2023.ccl-1.1
- Volume:
- Proceedings of the 22nd Chinese National Conference on Computational Linguistics
- Month:
- August
- Year:
- 2023
- Address:
- Harbin, China
- Editors:
- Maosong Sun, Bing Qin, Xipeng Qiu, Jing Jiang, Xianpei Han
- Venue:
- CCL
- SIG:
- Publisher:
- Chinese Information Processing Society of China
- Note:
- Pages:
- 1–16
- Language:
- Chinese
- URL:
- https://aclanthology.org/2023.ccl-1.1
- DOI:
- Cite (ACL):
- Jiayu Ding, Siyuan Wang, Zhongyu Wei, Qin Chen, and Xuanjing Huang. 2023. 基于推理链的多跳问答对抗攻击和对抗增强训练方法(Reasoning Chain Based Adversarial Attack and Adversarial Augmentation Training for Multi-hop Question Answering). In Proceedings of the 22nd Chinese National Conference on Computational Linguistics, pages 1–16, Harbin, China. Chinese Information Processing Society of China.
- Cite (Informal):
- 基于推理链的多跳问答对抗攻击和对抗增强训练方法(Reasoning Chain Based Adversarial Attack and Adversarial Augmentation Training for Multi-hop Question Answering) (Ding et al., CCL 2023)
- PDF:
- https://preview.aclanthology.org/nschneid-patch-4/2023.ccl-1.1.pdf