Tahir Ibrahim Gedi ¹ Tessfu Geteye Fantaye ^2*

• ¹ Jigjiga University, Jijiga, Somali Region, Ethiopia
• ² Dire Dawa University, Dire Dawa, Ethiopia

Morphological analysis is one of the vital steps in the development of several natural language processing applications. It decomposes words into their constituent morphemes. It is a challenging task to address the morphological features of morphologically rich languages like the Somali language. To overcome this challenge and advance previous studies on Somali morphological analysis, this study proposes a deep learning-based morphological analyzer with a larger corpus for the Somali language. Various recurrent neural network models with and without attention mechanisms have been developed and trained using a 19,839-word training dataset based on a 10-fold cross-validation technique. The trained models were evaluated using a separate 2,204-word testing dataset. The experimental results demonstrated that bidirectional long short-term memory (BILSTM) without attention, BILSTM with attention, bidirectional gated recurrent unit (BIGRU) without attention, and BIGRU with attention achieved accuracy values of 92.77%, 95.04%, 93.34%, and 95.92%, respectively. Based on these results, BIGRU with attention and BILSTM with attention models achieved the best accuracy for the Somali morphological analyzer. Overall, the recurrent neural network models with attention are suitable for developing the best-performing morphological analyzer for the Somali language.