Mining Term Translation from Domain Restricted Comparable Corpora

Several approaches have been proposed in the literature for extracting word translations from comparable corpora, almost all of them based on the idea of context similarity. This work addresses the aforementioned issue for the Basque-Spanish pair in a popular science domain. The main tasks our experiments focus on include: designing a method to combine some of the existing approaches; adapting this method to a popular science domain for the Basque-Spanish pair; and analyzing the performance of different approaches both for translating the contexts of the words and computing the similarity between contexts. We finally evaluate the different prototypes by calculating the precision for different cutoffs. The yielded results show the validity of the designed hybrid method, as well as the improvement obtained by using the probabilistic models we propose for computing the similarity between contexts.