Abstract
The recent surge of social media networks has provided a channel to gather and publish vital medical and health information. The focal role of these networks has become more prominent in periods of crisis, such as the recent pandemic of COVID-19. These social networks have been the leading platform for broadcasting health news updates, precaution instructions, and governmental procedures. They also provide an effective means for gathering public opinion and tracking breaking events and stories. To achieve location-based analysis for social media input, the location information of the users must be captured. Most of the time, this information is either missing or hidden. For some languages, such as Arabic, the users' location can be predicted from their dialects. The Arabic language has many local dialects for most Arab countries. Natural Language Processing (NLP) techniques have provided several approaches for dialect identification. The recent advanced language models using contextual-based word representations in the continuous domain, such as BERT models, have provided significant improvement for many NLP applications. In this work, we present our efforts to use BERT-based models to improve the dialect identification of Arabic text. We show the results of the developed models to recognize the source of the Arabic country, or the Arabic region, from Twitter data. Our results show 3.4% absolute enhancement in dialect identification accuracy on the regional level over the state-of-the-art result. When we excluded the Modern Standard Arabic (MSA) set, which is formal Arabic language, we achieved 3% absolute gain in accuracy between the three major Arabic dialects over the state-of-the-art level. Finally, we applied the developed models on a recently collected resource for COVID-19 Arabic tweets to recognize the source country from the users' tweets. We achieved a weighted average accuracy of 97.36%, which proposes a tool to be used by policymakers to support country-level disaster-related activities.