Statistical machine translation (SMT) refers to using probabilistic methods of learning translation process primarily from the parallel text. In SMT, the linguistic information such as morphology and syntax can be added to the parallel text for improved results. However, adding such linguistic matter is costly, in terms of time and expert effort. Here, we introduce a technique that can learn better shapes (morphological process) and more appropriate positioning (syntactic realization) of target words, without linguistic annotations. Our method improves result iteratively over multiple passes of translation. Our experiments showed better accuracy of translation, using a well-known scoring tool. There is no language specific step in this technique.