SITGs for Phrase Extraction and Mouse Actions in IMT

Abstract

This thesis presents two main contributions in the fields of Statistical Machine Translation and Interactive Machine Translation. In the field of Statistical Machine Translation, the efforts have been focused on obtaining high quality, linguistically motivated phrase pairs by means of Statistical Inversion Transduction Grammars. By using a SITG for parsing a bilingual corpus, spans are defined over both input and output strings, yielding the possibility of considering these spans as translations of each other. By doing so, phrase tables can be built from the bilingual corpus and fed to an off-the-shelf Statistical Machine Translation decoder. Moreover, novel syntax-based models are introduced in this thesis, and experimental results are shown which back up the inclusion of such models into the standard phrase translation table. Since these models are inherent to SITGs, they cannot be included into other standard phrase-based models. In the field of Interactive Machine Translation, a new interface between the user and the machine is proposed. By considering the Mouse Actions the user performs as an important input source for the system, it is shown that important and consistent performance gains may be achieved. These gains come in some cases at the cost of having the user ask for new suffix hypotheses, but in other cases these gains come at no cost, hence yielding true improvements to the state of the art.