Preliminary Examination Report: Bilingual Terminology Translation in Scientific Literature using Multilingual Structural Clues

In this report, we consider the problem of terminology translation in non-parallel scientific literature. We hypothesize that multilingual scientific literature has significant structure, both in the front matter and citations, that can be exploited for improved translation. After testing the well-known context vector method for translation from comparable corpora, we then propose an improvement which is able to exploit certain nearly-parallel segments of text present in the structure of multi-lingual scientific literature. Experimental results indicate that this method gives a substantial improvement in the ranking and accuracy of the system.