This work was presented at an international conference, The author proposed to use pre-trained models and context selection techniques to effectively encode the contextual information. The idea is to select the most significant context from existing data and encode contextual features using pre-trained models to have rich representation of context.

There can different context selection methodologies, we can have a number of previous sentences may have some relevant information about the topic and current sentence may have dependency on that. A current sentence that needs to be translated, a number of previous or next sentences can be selected as part of context. Our experiments suggest that considering more than 3 sentences as context may not improve the performance of system.