This thesis discusses how to incorporate linguistic knowledge into an SMT system. Although one important category of linguistic knowledge is that obtained by a constituent / dependency parser, a POS / super tagger, and a morphological analyser, linguistic knowledge here includes larger domains than this: Multi-Word Expressions, Out-Of-Vocabulary words, paraphrases, lexical semantics (or non-literal translations), named-entities, coreferences, and transliterations. The first discussion is about word alignment where we propose a MWE-sensitive word aligner. The second discussion is about the smoothing methods for a language model and a translation model where we propose a hierarchical Pitman-Yor process-based smoothing method. The common grounds for these discussion are the examination of three exceptional cases from real-world data: the presence of noise, the availability of prior knowledge, and the problem of underfitting. Notable characteristics of this design are the careful usage of (Bayesian) priors in order that it can capture both frequent and linguistically important phenomena. This can be considered to provide one example to solve the problems of statistical models which often aim to learn from frequent examples only, and often overlook less frequent but linguistically important phenomena.