Análise e comparação de técnicas de processamento de linguagem natural aplicadas à tradução automática de texto

Abstract

This work presents a comparative analysis between two prominent Neu ral Machine Translation (NMT) architectures: Recurrent Neural Networks (RNN) with GRU units and the Transformer architecture. The study starts from the hy pothesis that Transformers, due to their use of attention mechanisms, outperform recurrent approaches in terms of performance and generalization capacity. To test this hypothesis, two prototypes were implemented: an RNN-GRU model trained from scratch for English-to-French translation, which achieved 93.89% validation accuracy but showed limitations with unknown vocabulary and morphological com plexity; and a pre-trained Transformer model (mBART-50) for the same task. The results confirmed the superiority of the Transformer, which achieved higher per formance in metrics such as BLEU (0.8687) and METEOR (0.9329), generating fluent and accurate translations. The analysis concludes that the Transformer ar chitecture—widely recognized in the literature as the state of the art in machine translation (VASWANI et al., 2017)—offers a more robust and efficient solution for capturing the syntactic and semantic nuances of languages.