Have you often seen adversarial training used for sequence labeling to improve generalization across domains? The LSTM-CRF model appears to be the same model proposed by Lample et al (2016). I’d agree that it is common now to use that architecture.