Low resource word sense disambiguation in Oromo with fine tuned small transformers

Résumé

Abstract A key task in natural language processing is word sense disambiguation (WSD), which attempts to determine the accurate meaning of ambiguous words based on their context. While transformer-based designs have achieved significant results in high-resource languages, WSD for low-resource languages such as Oromo remains hard due to inadequate annotated corpora and lexical resources. Contextual representation learning has been greatly enhanced by recent advancements in transformer-based language models, allowing for more reliable disambiguation in situations with limited input. This study uses a manually created dataset from the Oromo–English Dictionary to examine the efficacy of transformer-based models for lexical-sample WSD in Oromo. The dataset contains sentences annotated by two native speakers, attaining an inter-annotator agreement of 0.82, indicating good annotation reliability. The dataset was filtered for experimental usage following preprocessing, normalization, and elimination of noisy cases. 472 training sentences, 71 validation sentences (15%), and 140 test sentences made up the final dataset. The dataset has a highly unbalanced long-tail distribution and encompasses 43 sense classes. BERT-base-cased gets the best performance with an accuracy of 0.862 and a macro-F1 score of 0.2897, according to an experimental evaluation of transformer-based models, including BERT, RoBERTa, DistilBERT, Davlan/afro-xlmr-base, and multilingual variations. Significant differences between models with χ 2 = 34.03 and p = 4.0 × 10 −1 are confirmed by statistical analysis using the Friedman test. BERT-base-cased performs much better than most transformer variations and classical baselines, according to post-hoc Wilcoxon signed-rank tests. These results show that contextual transformer representations are quite successful for low-resource WSD, although there is still a significant class imbalance that limits performance.

Citer ce document

Accès au document

Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter

Auteur(s)

Statistiques

Consultations : 1

Téléchargements : 0