Numerical Methods for Retrieval and Adaptation in Nagao’s EBMT model

We build an example-based machine translation system. It is an instance of case-based reasoning for machine translation. We introduce numerical methods instead of symbolic methods in two steps: retrieval and adaptation. For retrieval, we test three different approaches to define similarity between sentences. For adaptation, we use neural networks to solve analogies between sentences across languages. Oracle experiments allow to identify the best retrieval technique and to estimate the possibilities of such an approach. The system could place itself between a statistical and a neural machine translation systems on a task with not so large data.