<![CDATA[Text prediction is critical in natural language processing aiding as writing assistance. However, under-resourced languages lack support in current text prediction models. We examine research on text prediction standards towards modelling under-resourced languages. To achieve this, 806 studies were scrutinized, out of which 59 remained relevant. Key findings indicate the prevalence of N-gram, BERT, and LSTM models. A gap in the literature was noted, explaining why under-resourced languages are lagging. Precisely, data scarcity and domain specificity are the obstacles to progress in under-resourced language modeling. A potential solution to this challenge is visible, linked to the leverage of transfer learning techniques such as cross-lingual model pre-training. This way, data scarcity issues can be mitigated. We observed that N-gram models stand out in this respect. These are the most used text prediction approaches yielding outstanding perplexity scores. Our Shona text prediction prototype, RNN model achieved 83.69% accuracy with a perplexity score of 4.825. Notably, the N-gram based prototype outperformed the RNN model in all the measured categories. The development of text prediction standards will likely impact text prediction accuracy in under-resourced languages. Hopefully, research on under-resourced languages can draw insights from these standards and explore the development of tailored solutions.]]>
Copyrights © 2025
