Comparative Study of Word2Vec, FastText, and Glove Embeddings for Synonym Identification in Bugis Language

Abstract

This research aims to analyze the performance of word vectors generated by three pre-trained word embedding models, namely Word2Vec, Fast'Text, and Glove, to detect synonyms at the word level in the Bugis language. The word vectors from each model were integrated with a Bugis language dataset containing 12,060 monolingual texts, and then trained using the Long Short-Term Memory (LSTM) model. The results show that Word2Vec has the highest accuracy of 0.9412, followed by Glove with 0.9353 and FastText with 0.9262. Additional experiments with cosine similarity revealed that Glove performed best for predicting synonyms across all word occurrence frequency categories, while FastText and Word2Vec had inferior results. These findings indicate that although high training accuracy is achieved, it does not always imply that the model will excel in predicting synonyms.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint