A Comparison of Methods for Identifying the Translation of Words in a Comparable Corpus: Recipes and Limits

Abstract

Identifying translations in comparable corpora is a challenge that has attracted many researchers since a long time. It has applications in several applications including Machine Translation and Cross-lingual Information Retrieval. In this study we compare three state-of-the-art approaches for these tasks: the so-called context-based projection method, the projection of monolingual word embeddings, as well as a method dedicated to identify translations of rare words. We carefully explore the hyper-parameters of each method and measure their impact on the task of identifying the translation of English words in Wikipedia into French. Contrary to the standard practice, we designed a test case where we do not resort to heuristics in order to pre-select the target vocabulary among which to find translations, therefore pushing each method to its limit. We show that all the approaches we tested have a clear bias toward frequent words. In fact, the best approach we tested could identify the translation of a third of a set of frequent test words, while it could only translate around 10% of rare words.