LCUGA 2025 Presentation

This study examines the effectiveness of Transformer models in morphological segmentation for Nyangbo (Kwa, Gur), a tonal and endangered language spoken in Ghana. Four datasets—noisy, cleaned, noisy_reduced, and combined—were used to evaluate the performance of a supervised Transformer model. Results show that the model achieves superior performance with larger, noisy datasets and combined datasets compared to smaller, cleaned datasets. Specifically, segmentation and gloss prediction with noisy data resulted in an F1 score of 33.29%, outperforming the 29.96% F1 score achieved with cleaned data. For segmentation alone, the model recorded F1 scores of 62.57% (noisy) and 57.82% (cleaned), illustrating the importance of balancing data size and quality. Additionally, Morfessor 2.0 was trained in both supervised and unsupervised settings to evaluate its ability to process a low-resource tonal language, with the supervised model achieving an F1 score of 72%. Nyangbo’s unique linguistic features, including its agglutinative structure and complex tonal system, present significant challenges for morphological segmentation. In this language, entire sentences can be represented as single words. For example:

1. a-tɛ́-ba-dɛ́-mᴐ-ɛ́

   3SG-NEG-FUT-DIRECTIONAL-VERB-OBJ

   ‘S/he will not go and see her/him.

This example highlights the interaction of multiple morphemes, including the pronominal affix a-, negation affix tɛ ́ -, tense affix ba-, directional marker dɛ ́ -, verb mᴐ, and the object pronoun -í, which assimilates to -ɛ ́. The ability to accurately segment such structures is critical for computational linguistic applications.

Key findings suggest that combining noisy and cleaned datasets enhances generalization while using exclusively noisy data risks overfitting to noise patterns. Figures depicting training losses for noisy and cleaned datasets indicate that cleaner datasets promote better convergence, whereas noisy datasets introduce variability during training. Error analysis further revealed common challenges, including substitution errors driven by phonological similarities, omission of crucial morphemes, and unnecessary insertions, particularly with tonal and morphologically complex forms. These findings provide guidance for improving machine learning models for low-resource languages through strategic data preparation and curation.