The SLE example-based translation system.
Pete Whitelock, Victor Poznański · IWSLT · 2006
In this paper, we describe a hybrid approach to Machine Translation that exploits a corpus of example translations augmented with resources and techniques from rule-based MT. Our main motivation is to achieve reasonable accuracy for some subdomains with good time and space characteristics. Our architecture is relatively deterministic and therefore quite modest in its consumption of computational resources. At the same time, algorithms inspired by a view of translation in terms of string edits allow us to exploit some of the information available in the corpus to improve accuracy in a way that would be more difficult in other models. We describe the system which we developed at Sharp, illustrate how it exploits syntactic and semantic analysis for improved matching and disambiguation, and analyse our competition results.