Question answering (QA) systems have reached human-level accuracy; however, these systems are not robust enough and are vulnerable to adversarial examples. Recently, adversarial attacks have been widely investigated in text classification. However, there have been few research efforts on this topic in QA. In this article, we have modified the attack algorithms widely used in text classification to fit those algorithms for QA systems. We have evaluated the impact of various attack methods on QA systems at character, word, and sentence levels. Furthermore, we have developed a new framework, named RobustQA, as the first open-source toolkit for investigating textual adversarial attacks in QA systems. RobustQA consists of seven modules: Tokenizer, Victim Model, Goals, Metrics, Attacker, Attack Selector, and Evaluator. It currently supports six different attack algorithms. Furthermore, the framework simplifies the development of new attack algorithms in QA. The source code and documentation of RobustQA are available at https://github.com/mirbostani/RobustQA.

Named entity recognition (NER) is one of the major tasks in natural language processing. A named entity is often a word or expression that bears a valuable piece of information, which can be effectively employed by some major NLP tasks such as machine translation, question answering, and text summarization. In this paper, we introduce a new model called BERT-PersNER (BERT based Persian Named Entity Recognizer), in which we have applied transfer learning and active learning approaches to NER in Persian, which is regarded as a low-resource language. Like many others, we have used Conditional Random Field for tag decoding in our proposed architecture. BERT-PersNER has outperformed two available studies in Persian NER, in most cases of our experiments using the supervised learning approach on two Persian datasets called Arman and Peyma. Besides, as the very first effort to try active learning in the Persian NER, using only 30% of Arman and 20% of Peyma, we respectively achieved 92.15%, and 92.41% performance of the mentioned supervised learning experiments.