Towards an unsupervised morphological segmenter for isiXhosa

Abstract

In this paper, branching entropy techniques and isiXhosa language heuristics are adapted to develop unsupervised morphological segmenters for isiXhosa. An overview of isiXhosa segmentation issues is given, followed by a discussion on previous work in automated segmentation, and segmentation of isiXhosa in particular. Two unsupervised isiXhosa segmenters are presented and compared to a random minimum baseline and Morfessor-Baseline, a standard in unsupervised word segmentation. Morfessor-Baseline outperforms both isiXhosa segmenters at 79.10% boundary identification accuracy. The IsiXhosa Branching Entropy Segmenter (XBES) performance varies depending on the segmentation mode used, with a maximum of 73.39%. The IsiXhosa Heuristic Maximum Likelihood Segmenter (XHMLS) achieves 72.42%. The study suggests that unsupervised isiXhosa morphological segmentation is feasible with better optimization of the current attempt