Measuring semantic similarity between Arabic sentences

Abstract

This work discussed approaches of measuring Semantic Similarity between Arabic sentence, various techniques are described, then the results of similarity measurement using DSSM applied on the ASSD dataset was compared with results of HBMP model applied on the XNLI Arabic dataset. Another comparison is made between the results of the same HBMP model for the XNLI English and Arabic datasets. Our results show the higher accuracy achieved by HBMP on Arabic compared to DSSM, but this accuracy is still considered low compared to what the same model achieves for English, our conclusion focuses on the importance of a well curated dataset on the achievement of any model, especially in the context of Arabic which is morphologically complex and holds larger vocabulary set than English. Finally, it is suggested that Stemming can be utilized to enhance model accuracy for the purpose of semantic similarity between Arabic sentences.