Bilingual Word Embeddings from Parallel and Non-parallel Corpora for Cross-Language Text Classification