Summary of Deceiving Question-answering Models: a Hybrid Word-level Adversarial Approach, by Jiyao Li et al.
Deceiving Question-Answering Models: A Hybrid Word-Level Adversarial Approach
by Jiyao Li, Mingze Ni, Yongshun Gong, Wei Liu
First submitted to arxiv on: 12 Nov 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The proposed QA-Attack strategy is a novel word-level adversarial technique that can fool question answering (QA) models by creating deceptive inputs. By exploiting the customized attention mechanism and deletion ranking strategy, QA-Attack identifies and targets specific words within contextual passages, replacing them with synonyms to produce incorrect responses while preserving grammatical integrity. This approach demonstrates versatility across various question types and extensive long textual inputs. Experimental results show that QA-Attack successfully deceives baseline QA models and outperforms existing adversarial techniques in terms of success rate, semantic changes, BLEU score, fluency, and grammar error rate. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary QA-Attack is a new way to trick artificial intelligence language models into giving wrong answers. It works by changing certain words in a question or text to make the model think something that isn’t true. The strategy uses special attention mechanisms to figure out which words are most important and then replaces them with synonyms that won’t fool humans but will confuse AI models. This makes it possible to test how well these language models can withstand being tricked, and shows that QA-Attack is better than other methods at getting the models to make mistakes. |
Keywords
» Artificial intelligence » Attention » Bleu » Question answering
