Defending Against Disinformation Attacks in Open-Domain Question Answering

Explore a cutting-edge research presentation on defending against disinformation attacks in open-domain question answering systems. Delve into the innovative approach developed by researchers at the Center for Language & Speech Processing (CLSP) at Johns Hopkins University. Learn about their novel method that utilizes query augmentation to search for diverse passages, potentially less susceptible to poisoning, to answer original questions. Discover the concept of Confidence from Answer Redundancy (CAR), a new confidence method designed to integrate these passages into the model. Understand how this simple yet effective defense mechanism achieves significant improvements, with nearly 20% exact match gains across various levels of data poisoning and knowledge conflicts. Gain insights into the latest advancements in protecting open-domain question answering systems from adversarial attacks and enhancing their reliability in the face of disinformation.

Defending Against Disinformation Attacks in Open-Domain Question Answering - EACL 2024