An Efficient English Translation Practice Teaching System Using Gated Recurrent Unit with Attention Based Bidirectional Long Short Term Memory

Nowadays, as global communication intensifies English Translation Practice Teaching (ETPT) nurtures, cultural understanding, language mastery, and connectivity, preparing entities for productive association and exchange in multilingual environments. To enhance ETPT, many methods has been proposed earlier, which has overfitting, generalizability, and complexity problems. To overcome these problems, a model name called Gated Recurrent Unit with Attention based Bi-directional Long Short-Term Memory (GRU-Attention based Bi-LSTM) is proposed. A TED Talks Translation (TED-TT) dataset is considered for high quality translation. For preprocessing, tokenization, Part-Of-Speech (POS) tagging, named entity recognition, dependency parsing, stop-word removal, and stemming/lemmatization were considered. Word2Vec embedding is taken as feature extraction to handle out-of-vocabulary words, to reduce noise and outliers, and to increase generalizability. A five-fold cross-validation is considered and performed to handle imbalanced data, and to gain accurate performance. The classification involves transformer encoder with Bi-directional Encoder Representation Transformers (BERT) to handle sequential data and GRU-Attention based Bi-LSTM as classifier to handle long-term dependencies, to improve translation accuracy, and to reduce overfitting. The proposed GRU-Attention based Bi-LSTM model obtained higher performance metrics such as accuracy, precision, recall, and f1 score with 97.92%, 98.67%, 96.34%, and 94.86% when compared to existing Recurrent Neural Network (RNN).