Identification of Idioms by Machine Translation: a Hybrid Research System vs. Three Commercial Systems

We compare three commercial Machine Translation (MT) systems, Power Translator Pro, SYSTRAN, and T1 Langenscheidt, with the research hybrid, statistical and rule-based system, METIS-II, with respect to identification of idioms. Firstly, we make a distinction between continuous (adjacent constituents) and discontinuous idioms (non-adjacent constituents). Secondly, we describe our idiom resources within METIS-II, the system’s identification process, and we evaluate the results with simple techniques. From the translation outputs of the commercial systems we deduce that they cannot identify discontinuous idioms. We prove that, within METIS-II, the identification of discontinuous idioms is feasible, even with low resources.