Section: Software

MElt

Participants : Benoît Sagot [correspondant] , Pascal Denis.

MElt is a part-of-speech tagger, initially trained for French (on the French TreeBank and coupled with the Lefff), English [78] , Spanish, Kurmanji Kurdish [125] and Persian [106] , [107] . It is state-of-the-art for French. It is distributed freely as a part of the Alpage linguistic workbench.

In 2012, MElt has underwent two major upgrades:

• It has been successfully trained and used on Italian [35] , Spanish [26] and German data. In particular, a statistical parsing architecture for Italian that used MElt in a pre-processing step has obtained the best results in the EVALITA shared task on Italian parsing [35] .

• MElt can now be called within a wrapper developed for handling noisy textual data such as user-generated content produced on Web 2.0 platforms (forums, blogs, social media); more precisely, this wrapper is able to "clean" such data, then tag it using MElt, and finally transfer MElt annotations from the "cleaned" data, which could be annotated more easily, to the original noisy data. This architecture has proved useful on French for creating the French Social Media Bank [37] , [36] . On English, it has played an important role within both variants of the Alpage parsing architecture that were ranked 2nd and 3rd at the SANCL shared task on parsing user-generated content, organized by Google [38] .