Towards a Unified Approach to Memory- and Statistical-Based Machine Translation