Indonesian Word Sense Disambiguation Using KBBI Based Wordnet Contextual Embeddings
Abstract
Word Sense Disambiguation (WSD) is a fundamental challenge in Natural Language Processing (NLP), particularly for morphologically rich languages such as Indonesian. This paper presents a WSD system for Indonesian that leverages KBBI-Based WordNet synsets combined with a TF-IDF contextual embedding approach. We first construct Indonesian synsets automatically from Kamus Besar Bahasa Indonesia (KBBI) definitions using the DBSCAN clustering algorithm applied to TF-IDF cosine-similarity distance matrices, yielding 269 synsets covering 616 words across noun, adjective, and verb categories. The synsets are then integrated into a Modified Lesk WSD engine, where each sense definition is optionally augmented with synset gloss information to enrich semantic coverage. Disambiguation is performed by computing cosine similarity between a TF-IDF context vector and each candidate sense representation. Experimental results on a curated Indonesian test set of 20 ambiguous sentences demonstrate that the Synset-Enhanced TF-IDF Lesk model achieves an accuracy of 70.00%, precision of 72.22%, recall of 72.22%, and F1-score of 72.22%, substantially outperforming the random baseline of 40.00%. Per-word analysis reveals that disambiguation is most effective for nouns and adjectives with clearly differentiated KBBI definitions. These results establish a strong foundation for Indonesian WSD and highlight both the utility of KBBI-derived synsets and the limitations of purely lexicographic resources.