Minimal Data for Maximum Impact: An Indonesian Part-of-Speech Tagging Case Study

Annotating low-resource languages is challenging due to its time-consuming nature and high costs. In response to these challenges, this study investigates the potential of training models with minimal annotated data (520 tagged words) and abundant unannotated data (419,055 sentences), using Indonesian Part-of-Speech (POS) tagging as a case study. For the first time, we apply Stratos and Collins’ algorithm for Indonesian POS tagging, using four classifiers: Support Vector Machine, Naive Bayes, Decision Tree, and K-Nearest Neighbor. Our approach not only improved precision, recall, F1-score, and accuracy by approximately 1–5% compared to a baseline model that uses only the minimal annotated data, but also achieved a high performance for an Indonesian POS model, attaining an accuracy of 84% using this small amount of annotated data. This is a significant achievement in the context of using a small amount of annotated data to train the Indonesian POS model, as previous researchers have not achieved this level of accuracy with such a limited dataset. The method proved particularly beneficial for low-resource languages with limited high-quality annotated data but abundant unannotated data. It also reduced the workload of manual annotation as high-performance models required only a small amount of annotated data. Building on this efficiency, future work will focus on developing methods that further optimize the annotation process for low-resource language data.