跳到主要导航 跳到搜索 跳到主要内容

Evaluating and Enhancing the Robustness of Retrieval-Based Dialogue Systems with Adversarial Examples

  • Jia Li
  • , Chongyang Tao
  • , Nanyun Peng
  • , Wei Wu
  • , Dongyan Zhao
  • , Rui Yan*
  • *此作品的通讯作者
  • Peking University
  • University of Southern California
  • Microsoft USA

科研成果: 书/报告/会议事项章节会议稿件同行评审

摘要

Retrieval-based dialogue systems have shown strong performances on both consistency and fluency according to several recent studies. However, their robustness towards malicious attacks remains largely untested. In this paper, we generate adversarial examples in black-box settings to evaluate the robustness of retrieval-based dialogue systems. On three representative retrieval-based dialogue models, our attacks reduce R by 38.3 45.0 and 31.5 respectively on the Ubuntu dataset. Moreover, with adversarial training using our generated adversarial examples, we significantly improve the robustness of retrieval-based dialogue systems. We conduct thorough analysis to understand the robustness of retrieval-based dialog systems. Our results provide new insights to facilitate future work on building more robust dialogue systems.

源语言英语
主期刊名Natural Language Processing and Chinese Computing - 8th CCF International Conference, NLPCC 2019, Proceedings
编辑Jie Tang, Min-Yen Kan, Dongyan Zhao, Sujian Li, Hongying Zan
出版商Springer
142-154
页数13
ISBN(印刷版)9783030322328
DOI
出版状态已出版 - 2019
已对外发布
活动8th CCF International Conference on Natural Language Processing and Chinese Computing, NLPCC 2019 - Dunhuang, 中国
期限: 9 10月 201914 10月 2019

出版系列

姓名Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
11838 LNAI
ISSN(印刷版)0302-9743
ISSN(电子版)1611-3349

会议

会议8th CCF International Conference on Natural Language Processing and Chinese Computing, NLPCC 2019
国家/地区中国
Dunhuang
时期9/10/1914/10/19

指纹

探究 'Evaluating and Enhancing the Robustness of Retrieval-Based Dialogue Systems with Adversarial Examples' 的科研主题。它们共同构成独一无二的指纹。

引用此