There is a growing body of work in recent years to develop pre-trained language models (PLMs) for the Arabic language. This work addresses two major problems in existing Arabic PLMs that limit the progress of the Arabic NLU and NLG fields. First, existing Arabic PLMs are not well-explored and their pre-training can be improved significantly using a more methodical approach. Second, there is a lack of systematic and reproducible evaluation of these models in the literature. We revisit both the pre-training and evaluation of Arabic PLMs. In terms of pre-training, we explore the impact of the quality of the pretraining data, the size of the model, and the incorporation of character-level information on Arabic PLM. As a result, we release three new Arabic BERT-style models ( JABER, Char-JABER, and SABER), and two T5-style models (AT5S and AT5B). In terms of evaluation, we conduct a comprehensive empirical study to systematically evaluate the performance of existing state-of-the-art models on ALUE, a leaderboard-powered benchmark for Arabic NLU tasks, and on a subset of the Arabic generative tasks. We show that our models significantly outperform existing Arabic PLMs and achieve a new state-of-the-art performance on discriminative and generative Arabic NLU and NLG tasks. Our models and source code to reproduce results will be made available upon acceptance.

Abusive language in online discourse negatively affects a large number of social media users. Many computational methods have been proposed to address this issue of online abuse. The existing work, however, tends to focus on detecting the more explicit forms of abuse leaving the subtler forms of abuse largely untouched. Our work addresses this gap by making three core contributions. First, inspired by the theory of impoliteness, we propose a novel task of detecting a subtler form of abuse, namely unpalatable questions. Second, we publish a context-aware dataset for the task using data from a diverse set of Reddit communities. Third, we implement a wide array of learning models and also investigate the benefits of incorporating conversational context into computational models. Our results show that modeling subtle abuse is feasible but difficult due to the language involved being highly nuanced and context-sensitive. We hope that future research in the field will address such subtle forms of abuse since their harm currently passes unnoticed through existing detection systems.

Downstream effects of biased training data have become a major concern of the NLP community. How this may impact the automated curation and annotation of cultural heritage material is currently not well known. In this work, we create an experimental framework to measure the effects of different types of stylistic and social bias within training data for the purposes of literary classification, as one important subclass of cultural material. Because historical collections are often sparsely annotated, much like our knowledge of history is incomplete, researchers often cannot know the underlying distributions of different document types and their various sub-classes. This means that bias is likely to be an intrinsic feature of training data when it comes to cultural heritage material. Our aim in this study is to investigate which classification methods may help mitigate the effects of different types of bias within curated samples of training data. We find that machine learning techniques such as BERT or SVM are robust against reproducing the different kinds of bias within our test data, except in the most extreme cases. We hope that this work will spur further research into the potential effects of bias within training data for other cultural heritage material beyond the study of literature.