Evaluating Question Answering Model Robustness Against Typographical Mistakes Using Adversarial Attacks

• Ricardo Saraiva Grava USP
• Anarosa Alves Franco Brandão USP
• Sarajane Marques Peres USP
• Fabio Gagliardi Cozman USP

Resumo

Question Answering is one of the prime applications of LLMs and other such transformer-based language models, which have been shown to achieve results as of yet unmatched in many NLP tasks. And yet, the automatic evaluation metrics used in benchmarking these systems are unsatisfactory, producing results that don’t correlate well to human judgement and offer no interpretation of the model’s failings. This paper introduces Q3AE, an evaluation method for QA systems that leverages adversarial attacks to measure the robustness of answerer models when tested against challenging input. We test the method by introducing spelling mistakes and other such keyboarding errors, common in day to day use, into the SQuAD dataset, translated to Portuguese, and use it as input against LLMs of the LLaMA and Gemma architectures, of different sizes. We find that these models are surprisingly sensitive to even the tiniest alterations, responding poorly to modified examples that a human being could still easily understand.