Publication
Identifying main obstacles for statistical machine translation of morphologically rich South Slavic languages
Maja Popovic; Mihael Arčan
In: The Eighteenth Annual Conference of the European Association for Machine Translation (EAMT 15). Annual Conference of the European Association for Machine Translation (EAMT-15), May 11-13, Antalya, Turkey, Pages 97-104, EAMT, 2015.
Abstract
The best way to improve a statistical machine translation system is to identify concrete problems causing translation errors and address
them. Many of these problems are related to the characteristics of the involved languages and differences between them. This work explores
the main obstacles for statistical machine translation systems involving two morphologically rich and under-resourced languages, namely
Serbian and Slovenian. Systems are trained for translations from and into English and German using parallel texts from different
domains, including both written and spoken language. It is shown that for all translation directions structural properties concerning
multi-noun collocations and exact phrase boundaries are the most difficult for the systems, followed by negation, preposition and local word
order differences. For translation into English and German, articles and pronouns are the most problematic, as well as disambiguation of
certain frequent functional words. For translation into Serbian and Slovenian, cases and verb inflections are most difficult. In addition,
local word order involving verbs is often incorrect and verb parts are often missing, especially when translating from German.