EXPLORING LANGUAGE MODELS AND QUESTION ANSWERING IN BIOMEDICAL AND ARABIC DOMAINS

Abstract

Despite the success of the Transformer model and its variations (e.g., BERT, ALBERT, ELECTRA, T5) in addressing NLP tasks, similar success is not achieved when these models are applied to specific domains (e.g., biomedical) and limited-resources language (e.g., Arabic). This research addresses issues to overcome some challenges in the use of Transformer models to specialized domains and languages that lack in language processing resources.

One of the reasons for reduced performance in limited domains might be due to the lack of quality contextual representations. We address this issue by adapting different types of language models and introducing five BioM-Transformer models for the biomedical domain and Funnel transformer and T5 models for the Arabic language. For each of our models, we present experiments for studying the impact of design factors (e.g., corpora and vocabulary domain, model-scale, architecture design) on performance and efficiency. Our evaluation of BioM-Transformer models shows that we obtain state-of-the-art results on several biomedical NLP tasks and achieved the top-performing models on the BLURB leaderboard. The evaluation of our small scale Arabic Funnel and T5 models shows that we achieve comparable performance while utilizing less computation compared to the fine tuning cost of existing Arabic models. Further, our base-scale Arabic language models extend state-of-the-art results on several Arabic NLP tasks while maintaining a comparable fine-tuning cost to existing base-scale models.

Next, we focus on the question-answering task, specifically tackling issues in specialized domains and low-resource languages such as the limited size of question-answering datasets and limited topics coverage within them. We employ several methods to address these issues in the biomedical domain, including the employment of models adapted to the domain and Task-to-Task Transfer Learning. We evaluate the effectiveness of these methods at the BioASQ10 (2022) challenge, showing that we achieved the top-performing system on several batches of the BioASQ10 challenge. In Arabic, we address similar existing issues by introducing a novel approach to create question-answer-passage triplets, and propose a pipeline, Pair2Passage, to create large QA datasets. Using this method and the pipeline, we create the ArTrivia dataset, a new Arabic question-answering dataset comprising more than +10,000 high-quality question-answer-passage triplets. We presented a quantitative and qualitative analysis of ArTrivia that shows the importance of some often overlooked yet important components, such as answer normalization in enhancing the quality of the question-answer dataset and future annotation. In addition, our evaluation shows the ability of ArTrivia to build a question-answering model that can address the out-of-distribution issue in existing Arabic QA datasets.