Arabic Medical Community Question Answering Using ON-LSTM and CNN

In this paper, we address the problem of Arabic community question answering. We propose a model that leverages both the archived question and answer representations in the similarity computation with the user’s question. The proposed model considers the interaction of the user’s question with both archived questions and answers separately to address the noisy information problem in Arabic community question answering. The proposed model is a combination of two parts that covers question-question similarity and question-answer relevance. We used twin ON-LSTM with an attention mechanism and Arabic ELMo embeddings as input for the question-question similarity. For the question-answer relevance, we used a combination of twin ON-LSTM and CNN networks which can capture the relevance score even with long answers and questions. We evaluated the proposed model on the biomedical Arabic community question answering dataset cQA-MD. The proposed model outperformed the previous studies evaluated on the same dataset.