Comparative Analysis of LSTM and Bi-LSTM for Classifying Indonesian New Translation Bible Texts Using Word2Vec Embedding

The research aims to compare the classification accuracy of Long Short-Term Memory (LSTM) and Bidirectional LSTM (Bi-LSTM) architectures in classifying Indonesian New Translation Bible texts with Word2Vec embedding. The main objective is to examine how these deep learning models addressed complex and context-rich Indonesian biblical text classification, particularly between Old Testament and New Testament. The dataset used during the analysis contains 31,102 labeled verses with Word2Vec embedding calculated through both Continuous Bag of Words (CBOW) and Skip-Gram methods. Furthermore, the models are evaluated based on accuracy, precision, recall, and F1-score metrics. The results show Bi-LSTM performing better than LSTM in all cases, with the best performance being 92.31% when using Skip-Gram embedding at a vector size of 300 versus 91.94% by LSTM. The model demonstrates its ability to resolve semantic ambiguities in context-rich texts, such as the Bible. The research contributes to the text classification discipline through the provision of empirical evidence of the advantages of Bi-LSTM in managing biblical text processing as well as finding the optimal model architecture mix with word embedding methods. Additionally, the analysis also shows that future studies have to explore other embeddings, including GloVe (Global Vector) and FastText, or use transformer-based models such as Bidirectional Encoder Representations from Transformers (BERT) to improve the classification accuracy.