“…Similar as proposed in Gooding and Kochmar (2019); Hartmann and dos Santos (2018), and De Hertog and Tack (2018) we use word and character embeddings. We compare pretrained non-contextualized word embeddings, i.e., GloVe (Pennington et al, 2014), pre-trained contextualized word embeddings, i.e., ELMo (Peters et al, 2018) and BERT (Devlin et al, 2019), with pre-trained contextualized character embeddings, i.e., stacked Flair (Akbik et al, 2018(Akbik et al, , 2019a) -a combination of GloVe and Flair-and PooledFlair (Akbik et al, 2019b).…”