Deep Reinforcement Learning Method for Optimizing English Translation Algorithms

Abstract

Machine translation (MT) has revolutionized language processing, enabling real-time translation between languages with increasing accuracy. The challenge of optimizing English translation (ET) algorithms involves enhancing accuracy, fluency, and computational efficiency. Deep reinforcement learning (DRL) offers a promising solution, optimizing translation models by rewarding high-quality translations while minimizing errors. The purpose is to develop a DRL-based method, Gooseneck Barnacle-driven Enriched Double Deep Q Network (GB-EDDQN), to optimize ET algorithms, improving translation quality and efficiency. The dataset includes parallel corpora from bilingual/multilingual sources, providing sufficient training samples for ET. Text data undergo natural language processing (NLP) preprocessing, including tokenization, lemmatization, and stemming to standardize and simplify input sentences for better model performance, ensuring meaningful translations. Word2Vec is utilized for feature extraction, converting words into dense vector representations that capture semantic relationships. The proposed method, implemented in Python, achieves promising results with improved translation accuracy. Evaluation metrics in terms of recall (96.1%), accuracy (97.2%), and F1-score (95.4%) show a noticeable enhancement in translation fluency and coherence, demonstrating the effectiveness of GB-EDDQN in optimizing translation performance. By integrating reinforcement learning into the training process, the model adapts to translation challenges, enhancing both accuracy and fluency.