Abstract We consider Large-Scale Multi-Label Text Classifification (LMTC) in the legal domain. We release a new dataset of 57k legislative documents from EUR-LEX, annotated with ∼4.3k EUROVOC labels, which is suitable for LMTC, few- and zero-shot learning. Experimenting with several neural classififiers, we show that BIGRUs with label-wise attention perform better than other current state of the art methods. Domain-specifific WORD2VEC and context-sensitive ELMO embeddings further improve performance. We also fifind that considering only particular zones of the documents is suffificient. This allows us to bypass BERT’s maximum text length limit and fifinetune BERT, obtaining the best results in all but zero-shot learning cases