Section: New Results

Building a large-scale translation graph

Participants : Valérie Hanoka, Benoît Sagot.

Large-scale general-purpose multilingual translation databases are useful in a wide range of Natural Languages Processing (NLP) tasks. This is especially true concerning researches tackling problems specific to under-resourced languages, as translation databases can be used for adapting existing resources in other languages. This has been applied for example for the development of wordnets in languages other than English. There is thus a real need in NLP for open-source multilingual lexical databases that compiles as many translations as can be found on any freely available resource in any language.

We have developed, and are about to release, a new open-source heavily multilingual (over 590 languages) translation database built using several sources, namely various wiktionaries and the OPUS parallel corpora.

Our graph was built in several steps. We first extracted a preliminary set of translation and synonym pairs, which we stored in a large translation and synonym graph. We then applied filtering techniques for increasing the accuracy of this graph. We have evaluated the accuracy of our graph as being as high as 98% for translations extracted from wiktionaries.