Kernel-Based Machine Translation
Zhuoran Wang, John S. Shawe-Taylor · The MIT Press eBooks · 2008
This chapter presents a novel framework for machine translation based on kernel ridge regression. As a kernel method, the framework has the advantage of capturing the correspondences among the features of inputs and outputs in a very high-dimensional space. But the drawback is that its computational complexities are much higher than probabilistic models. A solution is sparse approximation, which poses the problem of extracting a sufficient amount of relevant bilingual training samples for a given input. Other essential improvements to this model could be the integration of additional language models and the utilization of linguistic knowledge.