Analisis Kinerja Word Embedding Glove dalam Penerjemahan Bahasa Batak-Inggris

Abstract

This study examines the performance of GloVe word embedding on the Long Short-Term Memory (LSTM) model in translating from Batak to English. GloVe has the ability to capture semantic meaning from a wide context of words. This study includes training the GloVe model with various parameters as well as collecting and processing a unique parallel Batak - English dataset. LSTM is a derivative of Recurrent Neural Network (RNN) which has the ability to maintain long-term dependencies and handle sequential data. In machine translation, LSTM performance is influenced by the quality of the word embedding used, which produces vector representations of words, and captures their semantic and contextual relationships. In this study, the authors analyze the performance of GloVe word embedding on the LSTM model and compare it with Word2Vec. The dataset used is 28,420 Batak-English sentence pairs collected from various sources. With encoder and decoder components, the LSTM model is trained for several epochs and the results are evaluated using the Bilingual Evalution Understudy (BLEU) score. This metric evaluates n-grams of actual translations with predicted translations, which then gives a translation accuracy score. The results show that GloVe word embedding performs better than Word2Vec. Glove word embedding gets an average BLEU score of 0.9415, while Word2Vec gets an average BLEU score of 0.9346. GloVe's better performance is due to its ability to understand language in a larger dataset and understand the context of words in a wider context