A Sentiment and Context-Aware Machine Translation Framework for Polite English-to-Tamil Conversion of Offensive Language with Cultural Sensitivity Consideration

Neural Machine Translation (NMT) involves the automated conversion of text from one language into another, aiming for linguistic accuracy and fluency.The domain of machine-based text translation and sentence matching undergoes rapid transformation due to swift developments in deep learning methods.However, training NMT systems demand extensive parallel datasets.For languages with scarce data resources, producing highquality translations remain a challenge, often resulting in less reliable outputs.Here, a politeness and sentiment-aware English-Tamil NMT module has been implemented.Initially, English text is gathered from the data sources and undergoes multitask classification with the developed Robustly Optimized Bidirectional Encoder Representations from Transformers Pretraining Approach-Residual Long Short-Term Memory (RoBERTa-ResLSTM) network, and the module finds the sentiment class and offensive class from the text.This module maps the English words into the compact word embedding space, and ResLSTM helps in finding the contextual semantics of the word.The resulting annotations related to the sentiment class and offensive class from the RoBERTa-ResLSTM are sent over to the politeness-controlled transfer Model.This module is built with the Multi-head Cross Attention-based mT5 (MC-AmT5) network.For better cross-handling of large multiple-language texts and sentence sequences, the proposed MC-AmT5 undergo hyperparameter tuning with the newly developed Modified Motion Vector-based Tornado optimizer with Coriolis force (MMV-MToC).The outcome from the politeness-controlled transfer model translated the English text into an equivalent Tamil text with better matching accuracy.For showcasing the effectiveness of the module, the model is experimented on standard datasets and compared against similar models.