Simplified Feature Set for Arabic Named Entity Recognition

This paper introduces simplified yet effective features that can robustly identify named entities in Arabic text without the need for morphological or syntactic analysis or gazetteers. A CRF sequence labeling model is trained on features that primarily use character n-gram of leading and trailing letters in words and word n-grams. The proposed features help overcome some of the morphological and orthographic complexities of Arabic. In comparing to results in the literature using Arabic specific features such POS tags on the same dataset and same CRF implementation, the results in this paper are lower by 2 F-measure points for locations, but are better by 8 points for organizations and 9 points for persons.