Detection of Adversarial Examples in Text Classification: Benchmark and Baseline via Robust Density Estimation
32514Authors
nan
Abstract
AbstractWord-level adversarial attacks have shown success in NLP models, drastically decreasing the performance of transformer-based models in recent years. As a countermeasure, adversarial defense has been explored, but relatively few efforts have been made to detect adversarial examples. However, detecting adversarial examples may be crucial for automated tasks (e.g. review sentiment analysis) that wish to amass information about a certain population and additionally be a step towards a robust defense system. To this end, we release a dataset for four popular attack methods on four datasets and four models to encourage further research in this field. Along with it, we propose a competitive baseline based on density estimation that has the highest auc on 29 out of 30 dataset-attack-model combinations. The source code is released (https://github.com/bangawayoo/adversarial-examples-in-text-classification).
Keywords
Full Content
Findings of the Association for Computational Linguistics: ACL 2022, pages 322 - 342 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Towards Transparent Interactive Semantic Parsing via Step-by-Step Correction Lingbo Mo, Ashley Lewis, Huan Sun, Michael White The Ohio State University {mo.169,lewis.2799,sun.397,white.1240}@osu.edu Abstract Existing studies on semantic parsing focus on mapping a natural-language utterance to a log- ical form (LF) in one turn. However, because natural language may contain ambiguity and variability, this is a difficult challenge. In this work, we investigate an interactive semantic parsing framework that explains the predicted LF step by step in natural language and enables the user to make corrections through natural- language feedback for individual steps. We focus on question answering over knowledge bases (KBQA) as an instantiation of our frame- work, aiming to increase the transparency of the parsing process and help the user trust the final answer. We construct INSPIRED, a crowdsourced dialogue dataset derived from the COMPLEXWEBQUESTIONS dataset. Our experiments show that this framework has the potential to greatly improve overall parse accu- racy. Furthermore, we develop a pipeline for dialogue simulation to evaluate our framework w.r.t. a variety of state-of-the-art KBQA models without further crowdsourcing effort. The re- sults demonstrate that our framework promises to be effective across such models.1 1 Introduction Semantic parsing aims to map natural language to formal meaning representations, such as λ-DCS, API calls, SQL and SPARQL queries. As seen in previous work (Liang et al., 2013; Yih et al., 2014, 2015; Talmor and Berant, 2018b; Chen et al., 2019; Lan and Jiang, 2020a; Gu et al., 2021), parsers still face major challenges: (1) the accuracy of state-of- the-art parsers is not high enough for real use, given that natural language questions can be ambiguous or highly variable with many possible paraphrases, and (2) it is hard for users to understand the parsing process and validate the results. In response to the challenges above, recent work (Li and Jagadish, 2014; He et al., 2016; 1Our INSPIRED dataset and code are available at https://github.com/molingbo/INSPIRED. What is the official language of the country that contains Al Sharqia Governorate? Not quite. Replace question 2 with "What is the official language spoken in the above-named nation?" Yep! Here's how I understood your question: 1. In what nation can you find the Al Sharqia Governorate? ANSWER: Egypt 2. What is the capital of the above-named nation? ANSWER: Cairo Is this what you were looking for? Sorry about that! How's this: 1. In what nation can you find the Al Sharqia Governorate? ANSWER: Egypt 2. What is the official language spoken in the above-named nation? ANSWER: Modern Standard Arabic Is this correct now? User Agent U1 A1 U2 U3 A2 Figure 1: Example dialogue from our dataset (dubbed INSPIRED). The agent turns (Ai’s) illustrate our em- phasis on transparency by explaining the predicted log- ical form step by step in natural language, along with intermediate answers, to the user for feedback. Chaurasia and Mooney, 2017; Su et al., 2018; Gur et al., …
Content truncated to 500 words
Labeled Text
AbstractWord-level adversarial attacks have shown success in NLP models, drastically decreasing the performance of transformer-based models in recent years. As a countermeasure, adversarial defense has been explored, but relatively few efforts have been made to detect adversarial examples. However, detecting adversarial examples may be crucial for automated tasks (e.g. review sentiment analysis) that wish to amass information about a certain population and additionally be a step towards a robust defense system. To this end, we release a dataset for four popular attack methods on four datasets and four models to encourage further research in this field. Along with it, we propose a competitive baseline based on density estimation that has the highest auc on 29 out of 30 dataset-attack-model combinations. The source code is released (
Publication Details
- Paper ID:
- 32514
- Year:
- 2022
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan