Addressing the Challenges of Igbo Computational Morphological Studies Using Frequent Pattern-Based Induction

Abstract

Computational studies of Igbo language are constrained by non-availability of large electronic corpora of Igbo text, a prerequisite for data-driven morphological induction. Existing unsupervised models, which are frequent-segment based, do not sufficiently address non-concatenative morphology and cascaded affixation prevalent in Igbo morphology, as well achieving affix labelling. This study devised a data-driven model that could induce non-concatenative aspects of Igbo morphology, cascaded affixation and affix labelling using frequent pattern-based induction. Tenfold Cross Validation (TCV) test was used to validate the propositions using percentages. An average accuracy measure of 88% was returned for the developed model. Ten purposively selected Igbo first speakers also evaluated samples of 100 model-analysed words each and the mean accuracy score of 82% was recorded. We conclude that morphology induction can be realized with a modestly sized corpus, demonstrating that electronic corpora scarcity does not constrain computational morphology studies as it would other higher levels of linguistic analysis.