Text classification systems have been proven vulnerable to adversarial text examples, modified versions of the original text examples that are often unnoticed by human eyes, yet can force text classification models to alter their classification. Often, research works quantifying the impact of adversarial text attacks have been applied only to models trained in English. In this paper, we introduce the first word-level study of adversarial attacks in Arabic. Specifically, we use a synonym (word-level) attack using a Masked Language Modeling (MLM) task with a BERT model in a black-box setting to assess the robustness of the state-of-the-art text classification models to adversarial attacks in Arabic. To evaluate the grammatical and semantic similarities of the newly produced adversarial examples using our synonym BERT-based attack, we invite four human evaluators to assess and compare the produced adversarial examples with their original examples. We also study the transferability of these newly produced Arabic adversarial examples to various models and investigate the effectiveness of defense mechanisms against these adversarial examples on the BERT models. We find that fine-tuned BERT models were more susceptible to our synonym attacks than the other Deep Neural Networks (DNN) models like WordCNN and WordLSTM we trained. We also find that fine-tuned BERT models were more susceptible to transferred attacks. We, lastly, find that fine-tuned BERT models successfully regain at least 2% in accuracy after applying adversarial training as an initial defense mechanism.

Wikipedia is a common source of training data for Natural Language Processing (NLP) research, especially as a source for corpora in languages other than English. However, for many downstream NLP tasks, it is important to understand the degree to which these corpora reflect representative contributions of native speakers. In particular, many entries in a given language may be translated from other languages or produced through other automated mechanisms. Language models built using corpora like Wikipedia can embed history, culture, bias, stereotypes, politics, and more, but it is important to understand whose views are actually being represented. In this paper, we present a case study focusing specifically on differences among the Arabic Wikipedia editions (Modern Standard Arabic, Egyptian, and Moroccan). In particular, we document issues in the Egyptian Arabic Wikipedia with automatic creation/generation and translation of content pages from English without human supervision. These issues could substantially affect the performance and accuracy of Large Language Models (LLMs) trained from these corpora, producing models that lack the cultural richness and meaningful representation of native speakers. Fortunately, the metadata maintained by Wikipedia provides visibility into these issues, but unfortunately, this is not the case for all corpora used to train LLMs.

The use of word embeddings is an important NLP technique for extracting meaningful conclusions from corpora of human text. One important question that has been raised about word embeddings is the degree of gender bias learned from corpora. Bolukbasi et al. (2016) proposed an important technique for quantifying gender bias in word embeddings that, at its heart, is lexically based and relies on sets of highly gendered word pairs (e.g., mother/father and madam/sir) and a list of professions words (e.g., doctor and nurse). In this paper, we document problems that arise with this method to quantify gender bias in diachronic corpora. Focusing on Arabic and Chinese corpora, in particular, we document clear changes in profession words used over time and, somewhat surprisingly, even changes in the simpler gendered defining set word pairs. We further document complications in languages such as Arabic, where many words are highly polysemous/homonymous, especially female professions words.

Natural Language Processing (NLP) systems are at the heart of many critical automated decision-making systems making crucial recommendations about our future world. Gender bias in NLP has been well studied in English, but has been less studied in other languages. In this paper, a team including speakers of 9 languages - Chinese, Spanish, English, Arabic, German, French, Farsi, Urdu, and Wolof - reports and analyzes measurements of gender bias in the Wikipedia corpora for these 9 languages. We develop extensions to profession-level and corpus-level gender bias metric calculations originally designed for English and apply them to 8 other languages, including languages that have grammatically gendered nouns including different feminine, masculine, and neuter profession words. We discuss future work that would benefit immensely from a computational linguistics perspective.