Research Article
Robustness Analysis on Natural Language Processing Based AI Q&A Robots
@INPROCEEDINGS{10.1007/978-3-030-32388-2_57, author={Chengxiang Yuan and Mingfu Xue and Lingling Zhang and Heyi Wu}, title={Robustness Analysis on Natural Language Processing Based AI Q\&A Robots}, proceedings={Machine Learning and Intelligent Communications. 4th International Conference, MLICOM 2019, Nanjing, China, August 24--25, 2019, Proceedings}, proceedings_a={MLICOM}, year={2019}, month={10}, keywords={AI security Question and answer robots Robustness Natural language processing}, doi={10.1007/978-3-030-32388-2_57} }
- Chengxiang Yuan
Mingfu Xue
Lingling Zhang
Heyi Wu
Year: 2019
Robustness Analysis on Natural Language Processing Based AI Q&A Robots
MLICOM
Springer
DOI: 10.1007/978-3-030-32388-2_57
Abstract
Recently, the natural language processing (NLP) based intelligent question and answering (Q&A) robots have been used in a wide range of applications, such as smart assistant, smart customer service, government business. However, the robustness and security issues of these NLP based artificial intelligence (AI) Q&A robots have not been studied yet. In this paper, we analyze the robustness problems in current Q&A robots, which include four aspects: (1) semantic slot settings are incomplete; (2) sensitive words are not filtered efficiently and completely; (3) Q&A robots return the search results directly; (4) unsatisfactory matching algorithms and inappropriate matching threshold settings. Then, we design and implement two types of evaluation tests, bad language and user’s typos, to evaluate the robustness of several state-of-the-art Q&A robots. Experiment results show that these common inputs (bad language and user’s typos) can successfully make these Q&A robots malfunction, denial of service, or speaking dirty words. Besides, we also propose possible countermeasures to enhance the robustness of these Q&A robots. To the best of the authors’ knowledge, this is the first work on analyzing the robustness and security problems of intelligent Q&A robots. This work can hopefully help provide guidelines to design robust and secure Q&A robots.