Investigating T5 Generation Neural Machine Translation Performance on English to German
Muhammad Raka Suryakusuma, Muhammad Faqih Ash Shiddiq, Henry Lucky, Irene Anindaputri Iswanto · 2023
This paper investigates the translation performance of the Text-to-Text Transfer Transformer (T5) model on English-to-German machine translation. The study aims to compare the performance of different generation methods, evaluate the quality of the dataset and tokenizer format, and explore the advantages and limitations of the T5 model. The evaluation metrics used are BLEU and BERTScore, which provide different scoring algorithms. The results show that the Greedy Inference method achieves the highest scores in both metrics, indicating its effectiveness in this task. The study contributes insights into the T5 model's performance for English-to-German translation and highlights the need for further improvements in lexical context and translation accuracy. Our experiments showed that the Greedy Inference method consistently achieved the highest scores in both BLEU and BERTScore metrics, indicating its effectiveness in generating translations for this task.