Study of NMT: An Explainable AI based approach

Machine translation (MT) which translates source to the target language is a challenging task. It is one of the important area under Natural language Processing (NLP).These days we are having neural based translation system viz. Neural Machine Translation (NMT). In this study we have discussed some of the important architecture of NMT systems and their working in brief. NMT uses deep neural network frame work in its design. Deep neural network based machine learning (ML) approach is entirely black box for end user. NMT requires several hyper parameter tuning during training. We still don't get any satisfactory answer of why we are getting any specific result (output) with specific set of linguistic features and the hyper parameters that are tuned. We have tried to focus on this aspect with very recent concept in AI i.e. explainable AI (XAI). XAI, can interpret and explain the internal behavior of model's architecture which can provide us reasons (explanation) when we get any result (output) from our model. We have discussed some of the techniques such as LIME, SHAPELY, Seq-2-Seq-Vis, LSTM-Vis which are widely used by data scientists and incorporate them in ML pipeline to design an unbiased, robust and accurate model.