Handling Out-of-Vocabulary in Indonesian POS Tagging: A Comparative Study
Muhammad Alfian, Umi Laili Yuhana, Daniel Fernando Siahaan, Harum Munazharoh, Eric Pardede · 2025
The out-of-vocabulary (OOV) problem is frequently encountered in POS tags. The OOV significantly affected the performance of the model in handling real-word situations. Several word embedding approaches have been proposed; however, no specific study has compared the role of word embedding in POS tagging in Indonesia. Therefore, this study investigated the effect of three word embeddings (Word2vec, GloVe, and Fasttext) on handling OOV in Indonesian. This study used a corpus from previous research comprising 355,021 words. This study divided the data using cross-validation and extracted orthographic and morphological information from each word. This study employs a state-of-theart POS tagging method, namely BiLSTM. The evaluation was performed in stages from training to testing with different levels of abstraction. The experimental results demonstrate that Fasttext outperforms other embedding methods with a performance increase of${1. 5 4 \%}$. The experimental results can be applied to other sequence labeling studies, such as NER.