A Comparative Study of Transformers Embeddings for Question Answering in Arabic Private Documents

Question answering (QA) tasks in natural language processing (NLP) are tricky, particularly when used with Arabic private documents. This is due to the complexity of Arabic language and the lack of sufficient annotated datasets for Arabic QA. Recent research findings have demonstrated the effectiveness of transformer embeddings for English QA. Their efficiency for QA in Arabic private documents, however, has not been fully investigated. In this study, various transformer embeddings for QA in Arabic private documents are compared. The effectiveness of these embeddings is evaluated by assessing a golden dataset extracted from private documents. The results show that the accuracy of QA in Arabic private documents can be significantly improved by combining transformer embeddings with count vectorization. Furthermore, different transformer embeddings have distinct strengths and weaknesses.