Decoding strategies for syntax-based statistical machine translation

Translation is the task of transforming text from a given language into another. Provided with a sentence in an input language, a human translator produces a sentence in the desired target language. The advances in artificial intelligence in the 1950s led to the idea of using machines instead of humans to generate translations. Based on this idea, the field of Machine Translation (MT) was created. The first MT systems aimed to map input text into the target translation through the application of hand-crafted rules. While this approach worked well for specific language-pairs on restricted fields, it was hardly extendable to new languages and domains because of the huge amount of human effort necessary to create new translation rules. The increase of computational power enabled Statistical Machine Translation (SMT) in the late 1980s, which addressed this problem by learning translation units automatically from large text collections. Statistical machine translation systems can be divided into several paradigms depending on the form of the (automatically learned) units used during translation. Early systems modeled translation between words. Later work extended these units from single words to sequences of words called phrases. A common point between word and phrase-based SMT is that the translation process takes place sequentially. This left-to-right process is not well suited to translate between languages where several words need to be reordered over (potentially) long distance. Such reorderings, which take place between many language pairs (e.g. English-German, English-Chinese or English-Arabic), led to the implementation of SMT systems based on formalisms that allow to translate recursively instead of sequentially. In these systems, called syntax-based systems, the (automatically learned) translation units are modeled with formal grammar productions and translation is performed by assembling the productions of these grammars. Many different grammar formalisms have been developed to model translation. One of the first is the Synchronous Context-Free Grammar (SCFG) which is an extension of the well-known Context-Free Grammar (CFG). To overcome several drawbacks of SCFG, more powerful formalisms have been explored such as the