Improving the precision of automatically constructed human-oriented translation dictionaries
Alexandra Antonova, Alexey Misyurev · 2014
In this paper we address the problem of automatic acquisition of a human-oriented translation dictionary from a large-scale parallel corpus. The initial translation equivalents can be extracted with the help of the techniques and tools developed for the phrase-table construction in statistical machine translation. The acquired transla-tion equivalents usually provide good lexi-con coverage, but they also contain a large amount of noise. We propose a super-vised learning algorithm for the detection of noisy translations, which takes into ac-count the context and syntax features, av-eraged over the sentences in which a given phrase pair occurred. Across nine Euro-pean language pairs the number of seri-ous translation errors is reduced by 43.2%, compared to a baseline which uses only phrase-level statistics. 1