On a Kernel Regression Approach to Machine Translation

摘要

We present a machine translation framework based on Kernel Regression techniques. The translation process is modeled as a string-to-string mapping. For doing so, first both source and target strings are mapped to a natural vector space obtaining feature vectors. Afterwards, a translation mapping is defined from the source feature vector to the target feature vector. This translation mapping, is learnt by linear regression. Once the target feature vector is obtained, we use a multi-graph search to find all the possible target strings whose mappings correspond to the "translated" feature vector. We present experiments in a small but relevant task showing encouraging results.