Dependency Structure Trees in Syntax Based Machine Translation

Machine Translation (MT) has long been an unsolved problem and more so an interesting and engaging problem. MT deals with the translation of a sentence in a source language into a sentence in the target language preserving the meaning in its full detail. This requires the computer to encode the knowledge of both languages in a representation that can be used at runtime to translate a give input. In earlier years translation problem was approached in three main ways. One, memorizing all the source and target sentences ahead of time called Translation Memory (TM) and reproducing the translation at runtime by a simple lookup process (Hutchins and Somers 1992). Second, use complete source language knowledge for analysis and generate the translation according to the syntax of the target language (Nirenburg et al. 1992). Third, a more semantically motivated and simultaneous multiple language oriented approach that projects the task of translation into a common space, with a uniform representation of knowledge called Interlingua (Hutchins and Somers 1992). Although these approaches have been experimented in great detail in the last two to three decades, a more promising approach called Statistical Machine Translation (SMT) with firm support from statistical and mathematical grounds has taken prominence in the last decade (Brown et al. 1993) (Koehn, Och, and Marcu 2003). Statistical Machine Translation (SMT) approaches use massive amounts of corpus to learn translation models at sub-sentential level which can generalize well to unseen data, unlike the TM. SMT addresses the problem of translation as a noisy channel paradigm where the channel model is usually called the ’Translation Model’ and the sourcemodel is called the ’LanguageModel’. The translationmodel is estimated under a generative story for word correspondences. The languagemodel estimated as ngram sequence models with markov assumptions. This formulation of the translation problem is language agnostic and does not assume any kind of syntax information from either the source or the target. Translation output produced under the SMT framework tend to usually be fragmented and context insensitive. With rigorous estimation techniques and heuristics for incorporation of context (Koehn, Och, andMarcu 2003), we have seen improved performance over the past few years. But, the quality is still far from human consistent. This is primarily to do with the fact that these models are learnt under no syntax scenarios and so are ill-informed about the phenomena of divergences that occur across languages. A recent body of approaches have looked into the incorporation of syntax at various phases of translation process with reasonable success (Yamada and Knight 2001; Chiang 2005). With more researchers looking into intelligent ways