Abstract
Most of the existing approaches to bilingual lexicon extraction (BLE) first map words in source and target languages into a single vector space, and then measure the similarity of words across the two languages in this space. We point out that existing BLE methods suffer from the so-called hubness phenomenon; i.e., a small number of translation candidates (hub candidates) are chosen by the systems as likely translations of many source words, which consequently degrade the accuracy of extracted translations. We show that this phenomenon can be alleviated by centering the data or by using the mutual proximity measure, which are two known techniques that effectively reduce hubness in standard nearest-neighbor search settings. Our empirical evaluation shows that naive nearest-neighbor search combined with these methods outperforms a recently proposed BLE method based on label propagation.
