This makes it consistent with the rest of the code and, especially, with the VocabularyDumper used to get fasttext embeddings, later queried by the LSTM+EMB classifier.
Coverage decreased (-0.04%) to 73.769% when pulling 5f8d32a0c3dc427335c8d2c26e04d1ed5650e68c on michelole:issue-86 into 2e4cd10e03b23076103abd4c4b4959159da2243f on bst-mug:master.
This makes it consistent with the rest of the code and, especially, with the
VocabularyDumper
used to get fasttext embeddings, later queried by the LSTM+EMB classifier.