Using Syllables as Features in Morpheme Tagging in Swahili

Utilizing corpora to build morphological analyzers for the purposes of computational application has been addressed in many different ways. Methods for automated morphological analysis generally focus on segmentation from raw text, and ignore the actual learning of what morpheme features are present. Other methods are time-consuming and require a great deal of prior knowledge of the language such as constructing a grammar by hand using finite-state transducers. We seek to create an analyzer which identifies which features are present without explicit segmentation and analysis. In this paper, we propose utilizing the surface-level cues of the morpheme based on the character sequences which generally comprise it as a guide for statistical morphological tag assignment in Bantu languages. In Swahili, these surface-level cues are syllabic in nature. This is grounded in typological insights from Bantu phonology; morphemes are generally monosyllabic, open syllables. Furthermore, this insight from Bantu is mirrored in current phonological theory. Optimality Theory (OT)(Prince and Smolensky, 2004) is a constraint-based approach at deriving a phonetic output from its phonological output, and vice versa. The phenomenon discussed above relates very closely to the OT constraint known as MORPH=σ1. This stipulates a typological phonological proclivity to relate morphemes to no more or less than a single syllable. The success of the approach we propose, therefore, should reaffirm not only the utility of such an insight, but its role in cur-